benchmark software peak performance for smart and brute force from node 0 for thesis

11 months ago · 317a74d164
66 changed files with 2888 additions and 35 deletions
--- a/benchmarks/benchmark-descriptors/modifier.py
+++ b/benchmarks/benchmark-descriptors/modifier.py
@ -10,41 +10,11 @@ def make_new_list(old,size):
    return new_list

 def process_json_file(file_path):
-    with open(file_path, 'r') as f:
-        json_data = json.load(f)
+    fr = open(file_path, 'r')
+    json_data = json.load(fr)
+        
+    json_data["path"] = "sw"

-        v_src = json_data["list"][0]["affinity"]["nnode_src"]
-        v_dst = json_data["list"][0]["affinity"]["nnode_dst"]
-
-        # Adjust v_src and v_dst based on HBM memory
-        v_src = v_src - 8 if v_src >= 8 else v_src
-        v_dst = v_dst - 8 if v_dst >= 8 else v_dst
-
-        if (0 <= v_src <= 3 and 0 <= v_dst <= 3):
-            list_size = 4
-            json_data["list"] = make_new_list(json_data["list"], list_size)
-            json_data["count"] = list_size
-            json_data["list"][0]["affinity"]["node"] = 0
-            json_data["list"][1]["affinity"]["node"] = 1
-            json_data["list"][2]["affinity"]["node"] = 2
-            json_data["list"][3]["affinity"]["node"] = 3
-        elif (4 <= v_src <= 7 and 4 <= v_dst <= 7):
-            list_size = 4
-            json_data["list"] = make_new_list(json_data["list"], list_size)
-            json_data["count"] = list_size
-            json_data["list"][0]["affinity"]["node"] = 4
-            json_data["list"][1]["affinity"]["node"] = 5
-            json_data["list"][2]["affinity"]["node"] = 6
-            json_data["list"][3]["affinity"]["node"] = 7
-        else:
-            # Case B: inter-socket operation
-            list_size = 2
-            json_data["list"] = make_new_list(json_data["list"], list_size)
-            json_data["count"] = list_size
-            json_data["list"][0]["affinity"]["node"] = v_src
-            json_data["list"][1]["affinity"]["node"] = v_dst
-
-    # Save modified JSON back to file
    with open(file_path, 'w') as f:
        json.dump(json_data, f, indent=2)

@ -56,5 +26,5 @@ def process_files_in_folder(folder_path):
                process_json_file(file_path)

 if __name__ == "__main__":
-    folder_path = "./benchmark-descriptors/peak-perf-smart/"
+    folder_path = "./peak-perf-allnodes-cpu/"
    process_files_in_folder(folder_path)
--- a/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton0-1gib-allnodes-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton0-1gib-allnodes-cpu.json
@ -0,0 +1,126 @@
+{
+  "count": 8,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 0,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 0,
+        "nnode_src": 0,
+        "node": 1
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 0,
+        "nnode_src": 0,
+        "node": 2
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 0,
+        "nnode_src": 0,
+        "node": 3
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 0,
+        "nnode_src": 0,
+        "node": 4
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 0,
+        "nnode_src": 0,
+        "node": 5
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 0,
+        "nnode_src": 0,
+        "node": 6
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 0,
+        "nnode_src": 0,
+        "node": 7
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "hw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton1-1gib-allnodes-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton1-1gib-allnodes-cpu.json
@ -0,0 +1,126 @@
+{
+  "count": 8,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 1,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 1,
+        "nnode_src": 0,
+        "node": 1
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 1,
+        "nnode_src": 0,
+        "node": 2
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 1,
+        "nnode_src": 0,
+        "node": 3
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 1,
+        "nnode_src": 0,
+        "node": 4
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 1,
+        "nnode_src": 0,
+        "node": 5
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 1,
+        "nnode_src": 0,
+        "node": 6
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 1,
+        "nnode_src": 0,
+        "node": 7
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "hw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton10-1gib-allnodes-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton10-1gib-allnodes-cpu.json
@ -0,0 +1,126 @@
+{
+  "count": 8,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 10,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 10,
+        "nnode_src": 0,
+        "node": 1
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 10,
+        "nnode_src": 0,
+        "node": 2
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 10,
+        "nnode_src": 0,
+        "node": 3
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 10,
+        "nnode_src": 0,
+        "node": 4
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 10,
+        "nnode_src": 0,
+        "node": 5
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 10,
+        "nnode_src": 0,
+        "node": 6
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 10,
+        "nnode_src": 0,
+        "node": 7
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "hw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton11-1gib-allnodes-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton11-1gib-allnodes-cpu.json
@ -0,0 +1,126 @@
+{
+  "count": 8,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 11,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 11,
+        "nnode_src": 0,
+        "node": 1
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 11,
+        "nnode_src": 0,
+        "node": 2
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 11,
+        "nnode_src": 0,
+        "node": 3
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 11,
+        "nnode_src": 0,
+        "node": 4
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 11,
+        "nnode_src": 0,
+        "node": 5
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 11,
+        "nnode_src": 0,
+        "node": 6
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 11,
+        "nnode_src": 0,
+        "node": 7
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "hw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton12-1gib-allnodes-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton12-1gib-allnodes-cpu.json
@ -0,0 +1,126 @@
+{
+  "count": 8,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 12,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 12,
+        "nnode_src": 0,
+        "node": 1
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 12,
+        "nnode_src": 0,
+        "node": 2
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 12,
+        "nnode_src": 0,
+        "node": 3
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 12,
+        "nnode_src": 0,
+        "node": 4
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 12,
+        "nnode_src": 0,
+        "node": 5
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 12,
+        "nnode_src": 0,
+        "node": 6
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 12,
+        "nnode_src": 0,
+        "node": 7
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "hw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton13-1gib-allnodes-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton13-1gib-allnodes-cpu.json
@ -0,0 +1,126 @@
+{
+  "count": 8,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 13,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 13,
+        "nnode_src": 0,
+        "node": 1
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 13,
+        "nnode_src": 0,
+        "node": 2
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 13,
+        "nnode_src": 0,
+        "node": 3
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 13,
+        "nnode_src": 0,
+        "node": 4
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 13,
+        "nnode_src": 0,
+        "node": 5
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 13,
+        "nnode_src": 0,
+        "node": 6
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 13,
+        "nnode_src": 0,
+        "node": 7
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "hw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton14-1gib-allnodes-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton14-1gib-allnodes-cpu.json
@ -0,0 +1,126 @@
+{
+  "count": 8,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 14,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 14,
+        "nnode_src": 0,
+        "node": 1
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 14,
+        "nnode_src": 0,
+        "node": 2
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 14,
+        "nnode_src": 0,
+        "node": 3
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 14,
+        "nnode_src": 0,
+        "node": 4
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 14,
+        "nnode_src": 0,
+        "node": 5
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 14,
+        "nnode_src": 0,
+        "node": 6
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 14,
+        "nnode_src": 0,
+        "node": 7
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "hw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton15-1gib-allnodes-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton15-1gib-allnodes-cpu.json
@ -0,0 +1,126 @@
+{
+  "count": 8,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 15,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 15,
+        "nnode_src": 0,
+        "node": 1
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 15,
+        "nnode_src": 0,
+        "node": 2
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 15,
+        "nnode_src": 0,
+        "node": 3
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 15,
+        "nnode_src": 0,
+        "node": 4
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 15,
+        "nnode_src": 0,
+        "node": 5
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 15,
+        "nnode_src": 0,
+        "node": 6
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 15,
+        "nnode_src": 0,
+        "node": 7
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "hw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton2-1gib-allnodes-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton2-1gib-allnodes-cpu.json
@ -0,0 +1,126 @@
+{
+  "count": 8,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 2,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 2,
+        "nnode_src": 0,
+        "node": 1
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 2,
+        "nnode_src": 0,
+        "node": 2
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 2,
+        "nnode_src": 0,
+        "node": 3
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 2,
+        "nnode_src": 0,
+        "node": 4
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 2,
+        "nnode_src": 0,
+        "node": 5
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 2,
+        "nnode_src": 0,
+        "node": 6
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 2,
+        "nnode_src": 0,
+        "node": 7
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "hw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton3-1gib-allnodes-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton3-1gib-allnodes-cpu.json
@ -0,0 +1,126 @@
+{
+  "count": 8,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 3,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 3,
+        "nnode_src": 0,
+        "node": 1
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 3,
+        "nnode_src": 0,
+        "node": 2
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 3,
+        "nnode_src": 0,
+        "node": 3
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 3,
+        "nnode_src": 0,
+        "node": 4
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 3,
+        "nnode_src": 0,
+        "node": 5
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 3,
+        "nnode_src": 0,
+        "node": 6
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 3,
+        "nnode_src": 0,
+        "node": 7
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "hw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton4-1gib-allnodes-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton4-1gib-allnodes-cpu.json
@ -0,0 +1,126 @@
+{
+  "count": 8,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 4,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 4,
+        "nnode_src": 0,
+        "node": 1
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 4,
+        "nnode_src": 0,
+        "node": 2
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 4,
+        "nnode_src": 0,
+        "node": 3
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 4,
+        "nnode_src": 0,
+        "node": 4
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 4,
+        "nnode_src": 0,
+        "node": 5
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 4,
+        "nnode_src": 0,
+        "node": 6
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 4,
+        "nnode_src": 0,
+        "node": 7
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "hw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton5-1gib-allnodes-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton5-1gib-allnodes-cpu.json
@ -0,0 +1,126 @@
+{
+  "count": 8,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 5,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 5,
+        "nnode_src": 0,
+        "node": 1
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 5,
+        "nnode_src": 0,
+        "node": 2
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 5,
+        "nnode_src": 0,
+        "node": 3
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 5,
+        "nnode_src": 0,
+        "node": 4
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 5,
+        "nnode_src": 0,
+        "node": 5
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 5,
+        "nnode_src": 0,
+        "node": 6
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 5,
+        "nnode_src": 0,
+        "node": 7
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "hw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton6-1gib-allnodes-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton6-1gib-allnodes-cpu.json
@ -0,0 +1,126 @@
+{
+  "count": 8,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 6,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 6,
+        "nnode_src": 0,
+        "node": 1
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 6,
+        "nnode_src": 0,
+        "node": 2
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 6,
+        "nnode_src": 0,
+        "node": 3
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 6,
+        "nnode_src": 0,
+        "node": 4
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 6,
+        "nnode_src": 0,
+        "node": 5
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 6,
+        "nnode_src": 0,
+        "node": 6
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 6,
+        "nnode_src": 0,
+        "node": 7
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "hw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton7-1gib-allnodes-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton7-1gib-allnodes-cpu.json
@ -0,0 +1,126 @@
+{
+  "count": 8,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 7,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 7,
+        "nnode_src": 0,
+        "node": 1
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 7,
+        "nnode_src": 0,
+        "node": 2
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 7,
+        "nnode_src": 0,
+        "node": 3
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 7,
+        "nnode_src": 0,
+        "node": 4
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 7,
+        "nnode_src": 0,
+        "node": 5
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 7,
+        "nnode_src": 0,
+        "node": 6
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 7,
+        "nnode_src": 0,
+        "node": 7
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "hw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton8-1gib-allnodes-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton8-1gib-allnodes-cpu.json
@ -0,0 +1,126 @@
+{
+  "count": 8,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 8,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 8,
+        "nnode_src": 0,
+        "node": 1
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 8,
+        "nnode_src": 0,
+        "node": 2
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 8,
+        "nnode_src": 0,
+        "node": 3
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 8,
+        "nnode_src": 0,
+        "node": 4
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 8,
+        "nnode_src": 0,
+        "node": 5
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 8,
+        "nnode_src": 0,
+        "node": 6
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 8,
+        "nnode_src": 0,
+        "node": 7
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "hw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton9-1gib-allnodes-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-allnodes-cpu/copy-n0ton9-1gib-allnodes-cpu.json
@ -0,0 +1,126 @@
+{
+  "count": 8,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 9,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 9,
+        "nnode_src": 0,
+        "node": 1
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 9,
+        "nnode_src": 0,
+        "node": 2
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 9,
+        "nnode_src": 0,
+        "node": 3
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 9,
+        "nnode_src": 0,
+        "node": 4
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 9,
+        "nnode_src": 0,
+        "node": 5
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 9,
+        "nnode_src": 0,
+        "node": 6
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 9,
+        "nnode_src": 0,
+        "node": 7
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "hw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton0-1gib-smart-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton0-1gib-smart-cpu.json
@ -0,0 +1,66 @@
+{
+  "count": 4,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 0,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 0,
+        "nnode_src": 0,
+        "node": 1
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 0,
+        "nnode_src": 0,
+        "node": 2
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 0,
+        "nnode_src": 0,
+        "node": 3
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "sw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton1-1gib-smart-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton1-1gib-smart-cpu.json
@ -0,0 +1,66 @@
+{
+  "count": 4,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 1,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 1,
+        "nnode_src": 0,
+        "node": 1
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 1,
+        "nnode_src": 0,
+        "node": 2
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 1,
+        "nnode_src": 0,
+        "node": 3
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "sw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton10-1gib-smart-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton10-1gib-smart-cpu.json
@ -0,0 +1,66 @@
+{
+  "count": 4,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 10,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 10,
+        "nnode_src": 0,
+        "node": 1
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 10,
+        "nnode_src": 0,
+        "node": 2
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 10,
+        "nnode_src": 0,
+        "node": 3
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "sw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton11-1gib-smart-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton11-1gib-smart-cpu.json
@ -0,0 +1,66 @@
+{
+  "count": 4,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 11,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 11,
+        "nnode_src": 0,
+        "node": 1
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 11,
+        "nnode_src": 0,
+        "node": 2
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 11,
+        "nnode_src": 0,
+        "node": 3
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "sw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton12-1gib-smart-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton12-1gib-smart-cpu.json
@ -0,0 +1,36 @@
+{
+  "count": 2,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 12,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 12,
+        "nnode_src": 0,
+        "node": 4
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "sw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton13-1gib-smart-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton13-1gib-smart-cpu.json
@ -0,0 +1,36 @@
+{
+  "count": 2,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 13,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 13,
+        "nnode_src": 0,
+        "node": 5
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "sw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton14-1gib-smart-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton14-1gib-smart-cpu.json
@ -0,0 +1,36 @@
+{
+  "count": 2,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 14,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 14,
+        "nnode_src": 0,
+        "node": 6
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "sw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton15-1gib-smart-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton15-1gib-smart-cpu.json
@ -0,0 +1,36 @@
+{
+  "count": 2,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 15,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 15,
+        "nnode_src": 0,
+        "node": 7
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "sw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton2-1gib-smart-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton2-1gib-smart-cpu.json
@ -0,0 +1,66 @@
+{
+  "count": 4,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 2,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 2,
+        "nnode_src": 0,
+        "node": 1
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 2,
+        "nnode_src": 0,
+        "node": 2
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 2,
+        "nnode_src": 0,
+        "node": 3
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "sw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton3-1gib-smart-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton3-1gib-smart-cpu.json
@ -0,0 +1,66 @@
+{
+  "count": 4,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 3,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 3,
+        "nnode_src": 0,
+        "node": 1
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 3,
+        "nnode_src": 0,
+        "node": 2
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 3,
+        "nnode_src": 0,
+        "node": 3
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "sw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton4-1gib-smart-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton4-1gib-smart-cpu.json
@ -0,0 +1,36 @@
+{
+  "count": 2,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 4,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 4,
+        "nnode_src": 0,
+        "node": 4
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "sw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton5-1gib-smart-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton5-1gib-smart-cpu.json
@ -0,0 +1,36 @@
+{
+  "count": 2,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 5,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 5,
+        "nnode_src": 0,
+        "node": 5
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "sw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton6-1gib-smart-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton6-1gib-smart-cpu.json
@ -0,0 +1,36 @@
+{
+  "count": 2,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 6,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 6,
+        "nnode_src": 0,
+        "node": 6
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "sw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton7-1gib-smart-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton7-1gib-smart-cpu.json
@ -0,0 +1,36 @@
+{
+  "count": 2,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 7,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 7,
+        "nnode_src": 0,
+        "node": 7
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "sw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton8-1gib-smart-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton8-1gib-smart-cpu.json
@ -0,0 +1,66 @@
+{
+  "count": 4,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 8,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 8,
+        "nnode_src": 0,
+        "node": 1
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 8,
+        "nnode_src": 0,
+        "node": 2
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 8,
+        "nnode_src": 0,
+        "node": 3
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "sw"
+}
--- a/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton9-1gib-smart-cpu.json
+++ b/benchmarks/benchmark-descriptors/peak-perf-smart-cpu/copy-n0ton9-1gib-smart-cpu.json
@ -0,0 +1,66 @@
+{
+  "count": 4,
+  "list": [
+    {
+      "affinity": {
+        "nnode_dst": 9,
+        "nnode_src": 0,
+        "node": 0
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 9,
+        "nnode_src": 0,
+        "node": 1
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 9,
+        "nnode_src": 0,
+        "node": 2
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    },
+    {
+      "affinity": {
+        "nnode_dst": 9,
+        "nnode_src": 0,
+        "node": 3
+      },
+      "task": {
+        "batching": {
+          "batch_size": 0,
+          "batch_submit": false
+        },
+        "iterations": 10,
+        "size": 1073741824
+      }
+    }
+  ],
+  "path": "sw"
+}
--- a/benchmarks/benchmark-descriptors/renamer.sh
+++ b/benchmarks/benchmark-descriptors/renamer.sh
@ -0,0 +1,19 @@
+#!/bin/bash
+
+json_directory="$1"
+
+# Check if the specified directory exists
+if [ ! -d "$json_directory" ]; then
+    echo "Error: Directory $json_directory not found. Provide as arg \$1"
+    exit 1
+fi
+
+# Iterate over all JSON files in the directory
+for json_file in "$json_directory"/*.json; do
+    # Check if there are matching files
+    if [ -e "$json_file" ]; then
+	bn=$(basename $json_file)
+	name="${bn%.*}"
+	mv $json_file "${json_directory}/${name}-cpu.json"
+    fi
+done
--- a/benchmarks/benchmark-results/copy-n0ton0-1gib-allnodes-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton0-1gib-allnodes-cpu-1e.json
@ -0,0 +1 @@
+{"count":8,"list":[{"affinity":{"nnode_dst":0,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[258881812,258316219,257971080,257844688,258663302,257832966,259203404,257843904,258622192,258546343],"completion":[258515198,257938981,257325262,257844164,258662842,257832435,259202883,257843472,258621764,258545911],"submission":[366614,377238,645818,524,460,531,521,432,428,432],"total":2325237361,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":0,"nnode_src":0,"node":1},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[262356597,262370375,261524117,262190702,262089819,262077325,262782165,262110669,261564411,249102821],"completion":[262354290,262017063,260836830,261538535,262089242,262076839,262781731,262110267,261564016,249102378],"submission":[2307,353312,687287,652167,577,486,434,402,395,443],"total":2345821564,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":0,"nnode_src":0,"node":2},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[270350913,269719654,269544161,269957879,269737472,270189457,269575365,270363367,270465593,202660522],"completion":[270347315,269296861,268914934,269322058,269736765,270188988,269574928,270362937,270465177,202660074],"submission":[3598,422793,629227,635821,707,469,437,430,416,448],"total":2362224577,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":0,"nnode_src":0,"node":3},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[271667243,271732421,272179264,272340723,272211710,272177041,272217483,272407257,272368705,167757678],"completion":[271665050,271426742,271525509,271657513,272211016,272176583,272217069,272406856,272368289,167757243],"submission":[2193,305679,653755,683210,694,458,414,401,416,435],"total":2345401260,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":0,"nnode_src":0,"node":4},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[624433635,308062288,275746264,276235706,275948413,276265188,275945154,275808422,275810592,275552034],"completion":[624431184,274653588,275576738,276085400,275947704,276264741,275944709,275808025,275810203,275551601],"submission":[2451,33408700,169526,150306,709,447,445,397,389,433],"total":2515479744,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":0,"nnode_src":0,"node":5},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[611258835,304646571,278982128,280501465,277661476,280752487,279319440,279474361,280131285,269419840],"completion":[611256420,279360346,278813154,280348341,277660773,280752016,279319028,279473953,280130873,269419360],"submission":[2415,25286225,168974,153124,703,471,412,408,412,480],"total":2530903576,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":0,"nnode_src":0,"node":6},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[560982046,301178762,292960517,293150375,294298537,293215813,291089630,293874502,292367318,221454475],"completion":[560980018,292522269,292812624,292983395,294297930,293215379,291089101,293874061,292366917,221454043],"submission":[2028,8656493,147893,166980,607,434,529,441,401,432],"total":2573606348,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":0,"nnode_src":0,"node":7},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[560595523,301772826,293313749,292570636,295657610,292550186,293021619,293216985,293992087,218486153],"completion":[560030732,293250130,293155868,292489192,295495712,292549590,293021201,293216560,293991667,218485710],"submission":[564791,8522696,157881,81444,161898,596,418,425,420,443],"total":2574596493,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"hw"}
--- a/benchmarks/benchmark-results/copy-n0ton0-1gib-smart-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton0-1gib-smart-cpu-1e.json
@ -0,0 +1 @@
+{"count":4,"list":[{"affinity":{"nnode_dst":0,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[250252769,249632814,249851745,249771929,249855535,249845039,249858780,249866685,249903445,249860682],"completion":[135,114,138,212,158,136,101,197,165,107],"submission":[250252634,249632700,249851607,249771717,249855377,249844903,249858679,249866488,249903280,249860575],"total":2248459900,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":0,"nnode_src":0,"node":1},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[265411230,265609643,265501635,265551498,265631372,265598306,265576330,265586974,265618509,258407365],"completion":[127,125,103,99,100,100,100,167,102,228],"submission":[265411103,265609518,265501532,265551399,265631272,265598206,265576230,265586807,265618407,258407137],"total":2383095025,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":0,"nnode_src":0,"node":2},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[273237103,273445323,273326440,273445660,273407829,273471806,273459803,273497105,271213634,262848027],"completion":[62,99,102,103,104,103,103,107,115,120],"submission":[273237041,273445224,273326338,273445557,273407725,273471703,273459700,273496998,271213519,262847907],"total":2448131016,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":0,"nnode_src":0,"node":3},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[286275910,286346276,286388173,286375724,286388515,286393525,286398847,286428730,278793384,269081148],"completion":[115,101,102,102,103,103,104,104,104,104],"submission":[286275795,286346175,286388071,286375622,286388412,286393422,286398743,286428626,278793280,269081044],"total":2552608793,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"sw"}
--- a/benchmarks/benchmark-results/copy-n0ton1-1gib-allnodes-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton1-1gib-allnodes-cpu-1e.json
@ -0,0 +1 @@
+{"count":8,"list":[{"affinity":{"nnode_dst":1,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[156787893,157468492,155633444,156415154,156123415,156152225,156248974,155997111,155852035,144105332],"completion":[156785104,155998168,155257187,156063892,156122822,156151760,156248473,155996697,155851582,144104874],"submission":[2789,1470324,376257,351262,593,465,501,414,453,458],"total":1394394038,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":1,"nnode_src":0,"node":1},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[153055594,152397268,152538028,153293305,152807971,153348998,153245728,153241431,152816362,152862522],"completion":[153051869,152184104,152174417,152931945,152807402,153348551,153245293,153240990,152815942,152862098],"submission":[3725,213164,363611,361360,569,447,435,441,420,424],"total":1376561785,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":1,"nnode_src":0,"node":2},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[163159925,164161443,163841764,164685868,164293062,164325101,164460840,163941774,163356111,96370999],"completion":[163157356,163978925,163473967,164311590,164292541,164324635,164460393,163941354,163355714,96370526],"submission":[2569,182518,367797,374278,521,466,447,420,397,473],"total":1409446136,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":1,"nnode_src":0,"node":3},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[161406200,162434508,162218886,162601257,162272474,162561753,162698050,161985169,162407633,109854990],"completion":[161403344,162252652,161840400,162253105,162271956,162561220,162697591,161984755,162407231,109854540],"submission":[2856,181856,378486,348152,518,533,459,414,402,450],"total":1409213191,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":1,"nnode_src":0,"node":4},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[364592446,235055390,215732756,216351998,216009895,215617425,215732727,216006934,217200414,203663656],"completion":[364589949,214966675,215609621,216234917,216009340,215616975,215732197,216006525,217200014,203663247],"submission":[2497,20088715,123135,117081,555,450,530,409,400,409],"total":1951384686,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":1,"nnode_src":0,"node":5},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[374298643,237335287,212046183,211573692,211191890,211586980,211314188,211417735,211730768,212112180],"completion":[374139833,208700231,211931083,211572914,211191476,211586231,211313603,211417299,211730320,212111756],"submission":[158810,28635056,115100,778,414,749,585,436,448,424],"total":1930397193,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":1,"nnode_src":0,"node":6},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[315939467,228530298,226164266,224938031,224590331,225034052,224092979,224022777,223873014,164547554],"completion":[315937190,225147758,226043879,224816074,224589759,225033610,224092421,224022339,223872606,164547116],"submission":[2277,3382540,120387,121957,572,442,558,438,408,438],"total":1965807789,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":1,"nnode_src":0,"node":7},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[326715235,226613642,222260268,222328956,222582952,223202760,223017975,222706470,222489731,178434084],"completion":[326392030,221961335,222142453,222266616,222461274,223202199,223017459,222706036,222489308,178433630],"submission":[323205,4652307,117815,62340,121678,561,516,434,423,454],"total":1963656024,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"hw"}
--- a/benchmarks/benchmark-results/copy-n0ton1-1gib-smart-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton1-1gib-smart-cpu-1e.json
@ -0,0 +1 @@
+{"count":4,"list":[{"affinity":{"nnode_dst":1,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[221753597,222151341,221955265,221733929,221498296,221746369,221752178,221766829,221597060,221143864],"completion":[107,112,123,166,104,104,105,110,206,111],"submission":[221753490,222151229,221955142,221733763,221498192,221746265,221752073,221766719,221596854,221143753],"total":1995381009,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":1,"nnode_src":0,"node":1},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[216979949,217013196,216639533,216702154,216606105,216598558,216800020,216684238,216702936,216727521],"completion":[143,293,107,39,35,37,37,39,45,83],"submission":[216979806,217012903,216639426,216702115,216606070,216598521,216799983,216684199,216702891,216727438],"total":1950488028,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":1,"nnode_src":0,"node":2},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[244653417,244757402,244709885,244656308,244760279,244770407,244627374,244409623,243577179,242173710],"completion":[103,102,101,100,100,207,100,156,183,143],"submission":[244653314,244757300,244709784,244656208,244760179,244770200,244627274,244409467,243576996,242173567],"total":2198466196,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":1,"nnode_src":0,"node":3},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[233076208,233116641,232963608,233000529,232946544,233097399,233052483,233040516,233195412,231369867],"completion":[105,106,103,104,151,102,102,101,211,106],"submission":[233076103,233116535,232963505,233000425,232946393,233097297,233052381,233040415,233195201,231369761],"total":2095802847,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"sw"}
--- a/benchmarks/benchmark-results/copy-n0ton10-1gib-allnodes-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton10-1gib-allnodes-cpu-1e.json
@ -0,0 +1 @@
+{"count":8,"list":[{"affinity":{"nnode_dst":10,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[95601895,95719985,95810402,95781376,95724529,95787290,95626317,95671619,95734339,95771054],"completion":[95599260,95618118,95646771,95618685,95723965,95786834,95625886,95671200,95733942,95770589],"submission":[2635,101867,163631,162691,564,456,431,419,397,465],"total":861708766,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":10,"nnode_src":0,"node":1},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[97214045,97228746,97290506,97391745,97209975,97233374,97284135,97294310,97230606,92212625],"completion":[97211872,97128552,97135724,97220386,97209432,97232899,97283716,97293903,97230192,92212182],"submission":[2173,100194,154782,171359,543,475,419,407,414,443],"total":870388205,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":10,"nnode_src":0,"node":2},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[98491028,98371595,98402716,98535940,98520496,98424806,98450036,98491779,98317119,84483032],"completion":[98426122,98287845,98255311,98535386,98520087,98424398,98449601,98491365,98316657,84482601],"submission":[64906,83750,147405,554,409,408,435,414,462,431],"total":872014017,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":10,"nnode_src":0,"node":3},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[99893885,99955455,100061649,100175514,99969749,99992484,99917047,100095384,100058003,71159943],"completion":[99891592,99853906,99902893,100029614,99969180,99992016,99916620,100094985,100057515,71159513],"submission":[2293,101549,158756,145900,569,468,427,399,488,430],"total":871435280,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":10,"nnode_src":0,"node":4},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[179378093,179313742,179653434,77822106,84050281,84133853,84111229,84105134,84173563,78739119],"completion":[179375456,179237254,179512078,77642045,84049573,84133321,84110835,84104712,84173177,78738591],"submission":[2637,76488,141356,180061,708,532,394,422,386,528],"total":936438709,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":10,"nnode_src":0,"node":5},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[181387228,181314717,181886076,76728960,85678065,85621617,85645172,85613608,85595829,68264324],"completion":[181237975,181188458,181743192,69922014,85603599,85620995,85644736,85613161,85595413,68263874],"submission":[149253,126259,142884,6806946,74466,622,436,447,416,450],"total":936445699,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":10,"nnode_src":0,"node":6},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[183226240,183206753,181240249,99865736,72233574,72226084,72202400,72270801,72188367,72234098],"completion":[183224056,183123551,181121931,72155891,72232220,72225604,72201973,72270373,72187974,72233695],"submission":[2184,83202,118318,27709845,1354,480,427,428,393,403],"total":897777064,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":10,"nnode_src":0,"node":7},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[183732992,184111001,179916869,102038883,76744703,76684116,76656850,76667508,76671885,66690513],"completion":[183731395,184039226,179773605,76715024,76743420,76683628,76656411,76667072,76671471,66690080],"submission":[1597,71775,143264,25323859,1283,488,439,436,414,433],"total":916238214,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"hw"}
--- a/benchmarks/benchmark-results/copy-n0ton10-1gib-smart-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton10-1gib-smart-cpu-1e.json
@ -0,0 +1 @@
+{"count":4,"list":[{"affinity":{"nnode_dst":10,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[137683699,137463203,137764970,137233736,137380315,137250388,137406322,137254876,137276647,136748390],"completion":[101,100,105,98,100,123,102,98,118,183],"submission":[137683598,137463103,137764865,137233638,137380215,137250265,137406220,137254778,137276529,136748207],"total":1235792307,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":10,"nnode_src":0,"node":1},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[145255506,145303763,145408176,145303322,145301868,145336728,145323770,145271775,145281557,143654599],"completion":[103,104,105,102,135,138,102,105,107,106],"submission":[145255403,145303659,145408071,145303220,145301733,145336590,145323668,145271670,145281450,143654493],"total":1306199528,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":10,"nnode_src":0,"node":2},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[134410839,134444366,134353692,133961608,133967172,133927591,133968134,133949578,133934015,134014008],"completion":[102,134,107,103,292,118,101,106,112,141],"submission":[134410737,134444232,134353585,133961505,133966880,133927473,133968033,133949472,133933903,134013867],"total":1206534295,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":10,"nnode_src":0,"node":3},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[143542649,143092920,143235379,143104338,143108249,143120144,143118753,143111202,143081433,141616979],"completion":[162,50,104,105,102,104,101,119,103,102],"submission":[143542487,143092870,143235275,143104233,143108147,143120040,143118652,143111083,143081330,141616877],"total":1286601395,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"sw"}
--- a/benchmarks/benchmark-results/copy-n0ton11-1gib-allnodes-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton11-1gib-allnodes-cpu-1e.json
@ -0,0 +1 @@
+{"count":8,"list":[{"affinity":{"nnode_dst":11,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[95800718,95947507,95733767,95735424,95769029,95845067,95787200,95653616,95871403,99944551],"completion":[95797920,95840962,95533507,95579307,95768479,95844629,95786768,95653206,95870995,99944126],"submission":[2798,106545,200260,156117,550,438,432,410,408,425],"total":866362179,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":11,"nnode_src":0,"node":1},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[97227819,97110752,97253860,97198697,97312544,97178159,97238916,97263705,97302989,92140692],"completion":[97225805,97016796,97110484,97057746,97311974,97177702,97238502,97263297,97302577,92140270],"submission":[2014,93956,143376,140951,570,457,414,408,412,422],"total":870010760,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":11,"nnode_src":0,"node":2},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[98383743,98486045,98523799,98751708,98564165,98501920,98632487,98443505,98463012,83983752],"completion":[98381495,98397871,98370493,98584120,98563616,98501454,98632041,98443048,98462565,83983315],"submission":[2248,88174,153306,167588,549,466,446,457,447,437],"total":872361004,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":11,"nnode_src":0,"node":3},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[99702663,99810369,99973477,99986468,99939764,99888486,99817023,99835516,99978213,72249802],"completion":[99700161,99710392,99812937,99828403,99939219,99888039,99816603,99835104,99977752,72249389],"submission":[2502,99977,160540,158065,545,447,420,412,461,413],"total":871488777,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":11,"nnode_src":0,"node":4},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[179685484,179429844,179919544,92029655,85622123,85666080,85601265,85600436,85574675,62325707],"completion":[179555634,179284685,179784096,88012975,85546963,85665366,85600875,85600017,85574283,62325272],"submission":[129850,145159,135448,4016680,75160,714,390,419,392,435],"total":941899951,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":11,"nnode_src":0,"node":5},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[180822615,181178446,179843373,94161746,84481885,84489885,84415210,84468165,84532124,58915487],"completion":[180820394,181099915,179713060,78240017,84480514,84489403,84414778,84467766,84531677,58915073],"submission":[2221,78531,130313,15921729,1371,482,432,399,447,414],"total":936646860,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":11,"nnode_src":0,"node":6},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[183185895,182867790,178064915,70932288,76244443,76158705,76202879,76209130,76170706,76207848],"completion":[183117125,182792914,177929952,70931522,76243947,76158273,76202439,76208695,76170294,76207416],"submission":[68770,74876,134963,766,496,432,440,435,412,432],"total":889128781,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":11,"nnode_src":0,"node":7},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[183921141,183992326,177268553,102316472,72543834,72412078,72507146,72477955,72470570,68821269],"completion":[183918719,183922267,177114482,72501532,72542564,72411606,72506686,72477430,72470163,68820839],"submission":[2422,70059,154071,29814940,1270,472,460,525,407,430],"total":894824679,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"hw"}
--- a/benchmarks/benchmark-results/copy-n0ton11-1gib-smart-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton11-1gib-smart-cpu-1e.json
@ -0,0 +1 @@
+{"count":4,"list":[{"affinity":{"nnode_dst":11,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[146175963,146224730,146233142,146186652,146206795,146164656,146273160,146202261,146179857,145112876],"completion":[106,105,108,105,104,127,105,101,103,101],"submission":[146175857,146224625,146233034,146186547,146206691,146164529,146273055,146202160,146179754,145112775],"total":1314797224,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":11,"nnode_src":0,"node":1},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[142536791,142491481,142505197,142515748,142540832,142529667,142486464,142495244,142536610,142280706],"completion":[39,44,38,38,42,43,39,39,39,64],"submission":[142536752,142491437,142505159,142515710,142540790,142529624,142486425,142495205,142536571,142280642],"total":1282394004,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":11,"nnode_src":0,"node":2},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[141368914,141437753,141414241,141393876,141408035,141374482,141407309,141419881,141366008,141234029],"completion":[60,103,101,103,102,101,103,102,103,111],"submission":[141368854,141437650,141414140,141393773,141407933,141374381,141407206,141419779,141365905,141233918],"total":1272471754,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":11,"nnode_src":0,"node":3},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[138709601,138709971,138686168,138720864,138722133,138735462,138697596,138704029,138664172,138694426],"completion":[101,101,106,103,102,104,100,139,103,101],"submission":[138709500,138709870,138686062,138720761,138722031,138735358,138697496,138703890,138664069,138694325],"total":1248345877,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"sw"}
--- a/benchmarks/benchmark-results/copy-n0ton12-1gib-allnodes-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton12-1gib-allnodes-cpu-1e.json
@ -0,0 +1 @@
+{"count":8,"list":[{"affinity":{"nnode_dst":12,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[167238545,167532623,167254279,163456519,135688238,93184157,93809914,93482621,93692225,93768161],"completion":[167235322,167409412,167027389,163240384,135687604,93183645,93809415,93482174,93691776,93767730],"submission":[3223,123211,226890,216135,634,512,499,447,449,431],"total":1102308599,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":12,"nnode_src":0,"node":1},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[170244393,170208157,170343124,166043196,127137873,96413209,96879163,96480334,96840369,88291565],"completion":[170241874,170093635,170123061,165829598,127137306,96412679,96878748,96479932,96839949,88291142],"submission":[2519,114522,220063,213598,567,530,415,402,420,423],"total":1108652029,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":12,"nnode_src":0,"node":2},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[182929300,182851818,183370460,172288928,110211185,110749847,110412188,110718229,79354311,56918774],"completion":[182926897,182730160,183154897,157093966,110209947,110749312,110411768,110717832,79353907,56918020],"submission":[2403,121658,215563,15194962,1238,535,420,397,404,754],"total":1116891462,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":12,"nnode_src":0,"node":3},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[183531814,183284995,183292201,159932705,109695534,110043070,109804601,110143322,83017962,54354147],"completion":[183308211,183073455,183069165,159932010,109695090,110042612,109804168,110142893,83017544,54353537],"submission":[223603,211540,223036,695,444,458,433,429,418,610],"total":1103586540,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":12,"nnode_src":0,"node":4},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[97009447,97305750,96957506,96851215,96953849,96969605,97332308,97051306,96948509,97090419],"completion":[97007325,97170614,96716497,96608065,96953145,96969142,97331881,96914066,96947828,97090001],"submission":[2122,135136,241009,243150,704,463,427,137240,681,418],"total":873596301,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":12,"nnode_src":0,"node":5},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[101762676,101861286,101622629,101874888,101443098,101702931,101452238,101893699,101814038,83748184],"completion":[101760537,101737130,101380480,101634567,101442580,101702491,101451655,101892649,101813613,83747763],"submission":[2139,124156,242149,240321,518,440,583,1050,425,421],"total":897424454,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":12,"nnode_src":0,"node":6},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[114709239,114498093,114788268,114923548,114915061,114492368,114865018,107863440,69757182,59632115],"completion":[114705922,114368871,114552808,114686215,114914529,114491921,114864327,107863035,69756610,59631657],"submission":[3317,129222,235460,237333,532,447,691,405,572,458],"total":925745741,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":12,"nnode_src":0,"node":7},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[114283075,114157340,114400990,114420056,114694550,114002217,114575352,108281612,67488610,56835263],"completion":[114279925,114028983,114168460,114196677,114693974,114001769,114574533,108281190,67488121,56834784],"submission":[3150,128357,232530,223379,576,448,819,422,489,479],"total":918870978,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"hw"}
--- a/benchmarks/benchmark-results/copy-n0ton12-1gib-smart-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton12-1gib-smart-cpu-1e.json
@ -0,0 +1 @@
+{"count":2,"list":[{"affinity":{"nnode_dst":12,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[231945400,231966647,231974244,231946021,231972874,231960486,231861730,233127720,233766353,233739910],"completion":[141,105,103,104,105,106,146,102,104,103],"submission":[231945259,231966542,231974141,231945917,231972769,231960380,231861584,233127618,233766249,233739807],"total":2092327674,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":12,"nnode_src":0,"node":4},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[185784726,185769268,185816689,185730251,185830706,185773704,185721487,185693492,185725292,185840893],"completion":[99,123,99,97,99,100,160,99,160,99],"submission":[185784627,185769145,185816590,185730154,185830607,185773604,185721327,185693393,185725132,185840794],"total":1671918184,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"sw"}
--- a/benchmarks/benchmark-results/copy-n0ton13-1gib-allnodes-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton13-1gib-allnodes-cpu-1e.json
@ -0,0 +1 @@
+{"count":8,"list":[{"affinity":{"nnode_dst":13,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[95870258,95765394,95882439,95684023,95981654,95930687,95886722,95893332,95917459,95821157],"completion":[95795376,95685593,95743439,95683439,95981206,95930236,95886297,95892892,95917009,95820741],"submission":[74882,79801,139000,584,448,451,425,440,450,416],"total":862773068,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":13,"nnode_src":0,"node":1},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[97181614,97240652,97306322,97331592,97221665,97189618,97253538,97366201,97199804,92931454],"completion":[97179201,97151541,97150971,97157515,97221105,97188951,97253084,97365763,97199388,92931020],"submission":[2413,89111,155351,174077,560,667,454,438,416,434],"total":871050026,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":13,"nnode_src":0,"node":2},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[98714450,98699324,98729893,98695644,98699586,98637674,98700174,98648480,98626766,83391598],"completion":[98544829,98529869,98544764,98616547,98548160,98637134,98699691,98648044,98626360,83391134],"submission":[169621,169455,185129,79097,151426,540,483,436,406,464],"total":872838545,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":13,"nnode_src":0,"node":3},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[99971782,100037585,100106968,100159287,100157567,100065859,99923351,100043428,100034413,72220458],"completion":[99969396,99955786,99939130,100009611,100157045,100065395,99922903,100042984,100034001,72220018],"submission":[2386,81799,167838,149676,522,464,448,444,412,440],"total":872758253,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":13,"nnode_src":0,"node":4},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[179562612,179674367,179809024,65071303,60449080,61031511,60956169,60999983,61019771,60965445],"completion":[179559966,179587131,179649949,64929008,60448481,61031037,60955592,60999558,61019332,60965016],"submission":[2646,87236,159075,142295,599,474,577,425,439,429],"total":789989104,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":13,"nnode_src":0,"node":5},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[181157026,181201511,181777031,61354745,63011866,62070482,62081654,62017144,62108008,59013733],"completion":[181154556,181120643,181611374,48593192,63010563,62069951,62081174,62016741,62107599,59013325],"submission":[2470,80868,165657,12761553,1303,531,480,403,409,408],"total":794650670,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":13,"nnode_src":0,"node":6},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[183448879,183540140,180545780,82931768,66562923,66460868,66617032,66632680,64401452,36221264],"completion":[183446729,183456187,180402810,62061310,66561687,66460357,66616615,66632209,64401050,36220793],"submission":[2150,83953,142970,20870458,1236,511,417,471,402,471],"total":813929037,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":13,"nnode_src":0,"node":7},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[184203475,184130040,178447153,106387848,66803666,66790336,66741461,66775549,52953342,34985069],"completion":[184201281,184043085,178293994,67724302,66802462,66789847,66741034,66775122,52952914,34984412],"submission":[2194,86955,153159,38663546,1204,489,427,427,428,657],"total":824030047,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"hw"}
--- a/benchmarks/benchmark-results/copy-n0ton13-1gib-smart-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton13-1gib-smart-cpu-1e.json
@ -0,0 +1 @@
+{"count":2,"list":[{"affinity":{"nnode_dst":13,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[225768947,226007442,226065827,225765702,225988185,225793064,225812530,226270409,227495329,227492887],"completion":[120,114,130,110,139,103,122,100,102,101],"submission":[225768827,226007328,226065697,225765592,225988046,225792961,225812408,226270309,227495227,227492786],"total":2036703408,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":13,"nnode_src":0,"node":5},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[189200798,189288784,189226699,189195824,189238103,189269988,189170444,189201572,189239050,189269941],"completion":[103,99,103,103,102,101,102,105,103,101],"submission":[189200695,189288685,189226596,189195721,189238001,189269887,189170342,189201467,189238947,189269840],"total":1703115599,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"sw"}
--- a/benchmarks/benchmark-results/copy-n0ton14-1gib-allnodes-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton14-1gib-allnodes-cpu-1e.json
@ -0,0 +1 @@
+{"count":8,"list":[{"affinity":{"nnode_dst":14,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[95657761,95596586,95654918,95663601,95529025,95730157,95712918,95580817,95624517,95625130],"completion":[95655251,95487447,95475746,95500458,95528483,95729672,95712471,95580380,95623876,95624723],"submission":[2510,109139,179172,163143,542,485,447,437,641,407],"total":860728473,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":14,"nnode_src":0,"node":1},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[97269978,97413947,97324294,97184312,97356308,97238375,97145295,97075950,97293420,92079747],"completion":[97092040,97218341,97167044,97098638,97172590,97237858,97144797,97075494,97292814,92079221],"submission":[177938,195606,157250,85674,183718,517,498,456,606,526],"total":870124417,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":14,"nnode_src":0,"node":2},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[98243104,98273265,98268762,98389776,98429222,98398731,98655794,98448072,98301522,84285511],"completion":[98240919,98180292,98121182,98199077,98428679,98398208,98655377,98370967,98300964,84285092],"submission":[2185,92973,147580,190699,543,523,417,77105,558,419],"total":871627246,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":14,"nnode_src":0,"node":3},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[99961887,99858989,99925605,100277866,99928430,99982470,99906950,100074619,100093815,71333663],"completion":[99959612,99751704,99778149,100078449,99927916,99981993,99906520,100073913,100093412,71333234],"submission":[2275,107285,147456,199417,514,477,430,706,403,429],"total":871393247,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":14,"nnode_src":0,"node":4},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[179610981,179783859,179839762,86157510,60968804,60909377,61056671,61061492,61017952,60990836],"completion":[179608383,179705377,179701904,63416194,60967419,60908701,61056252,61061099,61017534,60990425],"submission":[2598,78482,137858,22741316,1385,676,419,393,418,411],"total":811801411,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":14,"nnode_src":0,"node":5},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[181641632,181278833,178459980,75376092,62455644,62261343,62448364,62241223,62394665,60550726],"completion":[181639228,181203277,178328518,64745992,62454292,62260863,62447952,62240792,62394247,60550298],"submission":[2404,75556,131462,10630100,1352,480,412,431,418,428],"total":807481621,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":14,"nnode_src":0,"node":6},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[183737823,183425659,172505665,68986508,66073164,66234499,66034316,66207784,66108789,53069530],"completion":[183660474,183354044,172372401,68985791,66072713,66234071,66033872,66207349,66108378,53069098],"submission":[77349,71615,133264,717,451,428,444,435,411,432],"total":808736230,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":14,"nnode_src":0,"node":7},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[184184624,184621975,171002468,90769269,66685005,66706202,66652086,66750796,66682767,39679589],"completion":[184182174,184550427,170869199,68742202,66683795,66705721,66651630,66750378,66682346,39679136],"submission":[2450,71548,133269,22027067,1210,481,456,418,421,453],"total":819564393,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"hw"}
--- a/benchmarks/benchmark-results/copy-n0ton14-1gib-smart-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton14-1gib-smart-cpu-1e.json
@ -0,0 +1 @@
+{"count":2,"list":[{"affinity":{"nnode_dst":14,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[245237046,245277566,245321081,245301302,245261678,245270592,245191921,246533695,246744334,246737054],"completion":[101,101,100,104,100,102,114,101,108,104],"submission":[245236945,245277465,245320981,245301198,245261578,245270490,245191807,246533594,246744226,246736950],"total":2211651484,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":14,"nnode_src":0,"node":6},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[190149614,190093125,189958988,190012419,190070802,190155488,190220392,190130457,189992227,189539480],"completion":[104,102,104,102,103,102,105,103,100,188],"submission":[190149510,190093023,189958884,190012317,190070699,190155386,190220287,190130354,189992127,189539292],"total":1710190055,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"sw"}
--- a/benchmarks/benchmark-results/copy-n0ton15-1gib-allnodes-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton15-1gib-allnodes-cpu-1e.json
@ -0,0 +1 @@
+{"count":8,"list":[{"affinity":{"nnode_dst":15,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[95917736,95895747,95671689,95729529,96134841,95801671,95743640,95740993,95784252,95716146],"completion":[95784805,95732451,95498868,95653774,95963518,95799356,95743186,95740571,95783832,95715713],"submission":[132931,163296,172821,75755,171323,2315,454,422,420,433],"total":862614600,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":15,"nnode_src":0,"node":1},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[97212928,97158482,97386544,96971902,97046982,97209233,97124401,97250735,97225916,93011697],"completion":[97134290,97090286,97201400,96971374,97046459,97207952,97123975,97250300,97225504,93011267],"submission":[78638,68196,185144,528,523,1281,426,435,412,430],"total":870396743,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":15,"nnode_src":0,"node":2},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[98397748,98396088,98609268,98595098,98248332,98449178,98487717,98516423,98504594,84193181],"completion":[98395572,98309332,98445302,98432250,98247842,98448678,98487288,98516023,98504184,84192741],"submission":[2176,86756,163966,162848,490,500,429,400,410,440],"total":872010355,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":15,"nnode_src":0,"node":3},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[99819770,99816881,99930500,99961165,99714429,99888413,99918219,99983865,99920373,72489362],"completion":[99817615,99737022,99776808,99785918,99713922,99887899,99917814,99983439,99919968,72488962],"submission":[2155,79859,153692,175247,507,514,405,426,405,400],"total":871633171,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":15,"nnode_src":0,"node":4},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[179883849,179688644,179291145,60163191,54924856,60948644,60878959,60885553,60943555,60857568],"completion":[179881337,179612225,179161336,60024092,54924151,60948109,60878520,60885159,60943161,60857142],"submission":[2512,76419,129809,139099,705,535,439,394,394,426],"total":778595649,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":15,"nnode_src":0,"node":5},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[181188225,181397436,182303878,50219042,60899397,62118010,62259970,62113976,62070345,58982134],"completion":[181186142,181324342,182177563,39579971,60898077,62117502,62259446,62113565,62069929,58981705],"submission":[2083,73094,126315,10639071,1320,508,524,411,416,429],"total":782379675,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":15,"nnode_src":0,"node":6},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[183769217,183547044,179842499,106718239,66449537,66368584,66499758,66386953,49021731,34251198],"completion":[183766960,183468525,179701707,67385689,66448306,66368108,66499337,66386527,49021311,34250760],"submission":[2257,78519,140792,39332550,1231,476,421,426,420,438],"total":819101799,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":15,"nnode_src":0,"node":7},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[184080910,184160189,178424168,104531741,66663510,66684107,66641650,66723063,48610564,34787369],"completion":[184078842,184079818,178278557,67744804,66662206,66683646,66641209,66722640,48610175,34786941],"submission":[2068,80371,145611,36786937,1304,461,441,423,389,428],"total":817240375,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"hw"}
--- a/benchmarks/benchmark-results/copy-n0ton15-1gib-smart-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton15-1gib-smart-cpu-1e.json
@ -0,0 +1 @@
+{"count":2,"list":[{"affinity":{"nnode_dst":15,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[236591427,236569798,236722535,236690065,236716660,236692787,236625330,237445028,238194105,238253374],"completion":[104,103,104,104,106,132,103,103,102,104],"submission":[236591323,236569695,236722431,236689961,236716554,236692655,236625227,237444925,238194003,238253270],"total":2133923476,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":15,"nnode_src":0,"node":7},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[192030106,192022566,192038519,192099713,191926613,191880398,191911487,191911581,191995254,191897231],"completion":[138,49,46,53,47,49,46,48,48,51],"submission":[192029968,192022517,192038473,192099660,191926566,191880349,191911441,191911533,191995206,191897180],"total":1727698660,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"sw"}
--- a/benchmarks/benchmark-results/copy-n0ton2-1gib-allnodes-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton2-1gib-allnodes-cpu-1e.json
@ -0,0 +1 @@
+{"count":8,"list":[{"affinity":{"nnode_dst":2,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[162768547,162769391,159139012,163649555,162648613,162766457,162997447,162855491,162596464,113650138],"completion":[162766464,162586937,158793870,163284887,162648027,162765910,162996940,162855013,162595964,113649643],"submission":[2083,182454,345142,364668,586,547,507,478,500,495],"total":1413083381,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":2,"nnode_src":0,"node":1},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[164243665,164719299,160638192,165726531,164222094,164197582,164571145,164480797,164558062,100918709],"completion":[164241580,164531863,160263250,165370089,164221458,164197103,164570499,164480326,164557129,100918227],"submission":[2085,187436,374942,356442,636,479,646,471,933,482],"total":1414042592,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":2,"nnode_src":0,"node":2},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[154055112,153726171,153615778,149835443,153918873,153686088,154048565,153418451,153721406,153927129],"completion":[154052823,153506379,153258857,149470186,153918044,153685598,154047973,153418004,153720964,153926654],"submission":[2289,219792,356921,365257,829,490,592,447,442,475],"total":1379907212,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":2,"nnode_src":0,"node":3},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[156841281,156968922,156535914,168078683,156565983,156642546,157363035,156287362,156304296,142101295],"completion":[156839058,156762595,156180089,156790377,156564634,156641886,157362510,156286876,156303764,142100802],"submission":[2223,206327,355825,11288306,1349,660,525,486,532,493],"total":1407044536,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":2,"nnode_src":0,"node":4},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[317861900,222814462,223614852,223700094,224045344,223392853,225483471,222748609,224700610,175055684],"completion":[317723410,222754906,223501384,223699555,224044894,223392427,225483026,222748202,224700197,175055251],"submission":[138490,59556,113468,539,450,426,445,407,413,433],"total":1965568239,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":2,"nnode_src":0,"node":5},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[312929579,224657996,224631870,225262655,224273491,225444618,225889882,223611937,225564847,162180091],"completion":[312630777,224540746,224515380,225204778,224154687,225444008,225889461,223611524,225564435,162179629],"submission":[298802,117250,116490,57877,118804,610,421,413,412,462],"total":1961531297,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":2,"nnode_src":0,"node":6},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[360019967,233041787,212846600,213759074,213360474,213750218,212783152,213543629,213325531,213048077],"completion":[360017609,213700723,212730541,213638974,213359938,213749728,212782720,213543230,213325139,213047647],"submission":[2358,19341064,116059,120100,536,490,432,399,392,430],"total":1939472919,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":2,"nnode_src":0,"node":7},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[350811405,227777219,216014667,215552541,216066129,216287238,216322146,216664466,216280065,208824804],"completion":[350808797,215763314,215897297,215434633,216065498,216286770,216321718,216664065,216279608,208824360],"submission":[2608,12013905,117370,117908,631,468,428,401,457,444],"total":1949802653,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"hw"}
--- a/benchmarks/benchmark-results/copy-n0ton2-1gib-smart-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton2-1gib-smart-cpu-1e.json
@ -0,0 +1 @@
+{"count":4,"list":[{"affinity":{"nnode_dst":2,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[245429775,245477448,245430102,245471203,245487372,245444762,245423450,245437784,245415055,245369019],"completion":[99,98,96,18,239,115,94,195,99,105],"submission":[245429676,245477350,245430006,245471185,245487133,245444647,245423356,245437589,245414956,245368914],"total":2208968494,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":2,"nnode_src":0,"node":1},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[255230894,255252408,255200214,255284944,255229783,255307558,255171803,255166665,255242919,254431117],"completion":[98,38,107,104,104,103,103,104,103,102],"submission":[255230796,255252370,255200107,255284840,255229679,255307455,255171700,255166561,255242816,254431015],"total":2296301646,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":2,"nnode_src":0,"node":2},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[243338472,243277058,243274919,243350350,243339668,243280316,243305705,243260308,243287616,243189979],"completion":[97,110,101,113,151,212,230,98,104,99],"submission":[243338375,243276948,243274818,243350237,243339517,243280104,243305475,243260210,243287512,243189880],"total":2189579858,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":2,"nnode_src":0,"node":3},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[255665848,255687421,255698760,255768383,255744534,255765617,255712683,255712747,255782402,254920320],"completion":[102,99,97,101,50,51,50,50,51,94],"submission":[255665746,255687322,255698663,255768282,255744484,255765566,255712633,255712697,255782351,254920226],"total":2300805138,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"sw"}
--- a/benchmarks/benchmark-results/copy-n0ton3-1gib-allnodes-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton3-1gib-allnodes-cpu-1e.json
@ -0,0 +1 @@
+{"count":8,"list":[{"affinity":{"nnode_dst":3,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[163803221,163850568,164120545,163959909,164209404,164090484,163858569,164353426,163495083,103738927],"completion":[163800543,163656745,163763344,163600921,164208833,164090052,163858104,164352986,163494656,103738269],"submission":[2678,193823,357201,358988,571,432,465,440,427,658],"total":1415687834,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":3,"nnode_src":0,"node":1},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[162159606,161493023,161720140,161832370,162495675,161863178,161439497,161880990,161349922,119896830],"completion":[162157563,161300314,161326203,161478641,162495055,161862742,161439098,161880549,161349505,119896400],"submission":[2043,192709,393937,353729,620,436,399,441,417,430],"total":1414385778,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":3,"nnode_src":0,"node":2},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[157446650,157774060,156903795,157355427,156834496,157626025,157009546,157660808,157378124,144800360],"completion":[157444276,157533746,156535987,156959855,156833844,157625496,157009113,157660378,157377717,144799943],"submission":[2374,240314,367808,395572,652,529,433,430,407,417],"total":1403352340,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":3,"nnode_src":0,"node":3},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[154272430,154318458,154702589,153988190,153986896,154027426,154000327,154207114,154079404,153031512],"completion":[154269221,154116666,154346555,153605721,153986239,154026993,153999880,154206664,154079010,153031086],"submission":[3209,201792,356034,382469,657,433,447,450,394,426],"total":1386351009,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":3,"nnode_src":0,"node":4},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[317846025,226766586,225427116,225483797,225924368,225409208,225660272,224518129,225876126,162637565],"completion":[317534555,224626315,225308111,225483230,225923962,225408768,225659841,224517731,225875692,162637148],"submission":[311470,2140271,119005,567,406,440,431,398,434,417],"total":1967718093,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":3,"nnode_src":0,"node":5},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[335256226,227897713,223784313,223848716,224323653,223690775,224495855,222331523,222900114,173469830],"completion":[335254118,223686134,223667685,223726391,224323081,223690321,224495435,222331068,222899691,173469413],"submission":[2108,4211579,116628,122325,572,454,420,455,423,417],"total":1966756414,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":3,"nnode_src":0,"node":6},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[354240483,229035152,216434880,216094051,215247678,215772508,215702743,215865604,216031152,205560216],"completion":[354238376,214729344,216311862,215969261,215247095,215772050,215702262,215865192,216030713,205559686],"submission":[2107,14305808,123018,124790,583,458,481,412,439,530],"total":1945757788,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":3,"nnode_src":0,"node":7},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[372539890,237037624,211005256,211170971,211594933,211676547,211411693,211911916,211296007,211040517],"completion":[372536849,210508383,210887668,211053495,211594386,211676036,211411255,211911509,211295608,211040111],"submission":[3041,26529241,117588,117476,547,511,438,407,399,406],"total":1928158594,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"hw"}
--- a/benchmarks/benchmark-results/copy-n0ton3-1gib-smart-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton3-1gib-smart-cpu-1e.json
@ -0,0 +1 @@
+{"count":4,"list":[{"affinity":{"nnode_dst":3,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[232716078,232763602,232642125,232748551,232608318,232679489,232654743,232637738,232690025,232289984],"completion":[109,100,196,101,116,107,99,103,105,107],"submission":[232715969,232763502,232641929,232748450,232608202,232679382,232654644,232637635,232689920,232289877],"total":2093726991,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":3,"nnode_src":0,"node":1},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[228159635,228262283,228122520,228213151,228070918,228177523,228103195,228082363,228087545,228178568],"completion":[105,135,104,103,110,104,101,103,103,103],"submission":[228159530,228262148,228122416,228213048,228070808,228177419,228103094,228082260,228087442,228178465],"total":2053309814,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":3,"nnode_src":0,"node":2},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[227692726,227725378,227657988,227650628,227634794,227669122,227614136,227558245,227634061,227570766],"completion":[102,104,106,104,102,98,102,103,102,105],"submission":[227692624,227725274,227657882,227650524,227634692,227669024,227614034,227558142,227633959,227570661],"total":2048727893,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":3,"nnode_src":0,"node":3},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[228046177,228051928,228093322,227952563,228020875,227993480,228029166,227949776,227950131,228067276],"completion":[103,104,113,110,103,117,165,117,98,151],"submission":[228046074,228051824,228093209,227952453,228020772,227993363,228029001,227949659,227950033,228067125],"total":2052122380,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"sw"}
--- a/benchmarks/benchmark-results/copy-n0ton4-1gib-allnodes-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton4-1gib-allnodes-cpu-1e.json
@ -0,0 +1 @@
+{"count":8,"list":[{"affinity":{"nnode_dst":4,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[361417398,360369348,305292137,199867954,207575432,207486729,207853351,208485182,207640229,208119075],"completion":[361243892,360204344,304982142,199867268,207574917,207486218,207852898,208484703,207639795,208118583],"submission":[173506,165004,309995,686,515,511,453,479,434,492],"total":2112838642,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":4,"nnode_src":0,"node":1},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[363418556,364867832,297511419,232597794,213387519,212096904,212184393,211760603,212225238,192629768],"completion":[363416442,364690921,297191735,211862266,213386113,212096438,212183962,211760123,212224785,192629330],"submission":[2114,176911,319684,20735528,1406,466,431,480,453,438],"total":2149277105,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":4,"nnode_src":0,"node":2},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[378170967,381411922,259135122,221411767,221196003,220809414,221437631,220556697,221051773,149373667],"completion":[378167976,381231663,258779048,221285532,221195466,220808973,221437232,220556250,221051300,149373246],"submission":[2991,180259,356074,126235,537,441,399,447,473,421],"total":2116396022,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":4,"nnode_src":0,"node":3},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[382303433,376764799,287839387,223214456,222937874,223061201,223307562,222856726,221092469,113052263],"completion":[382300911,376593762,248036595,223072157,222937293,223060760,223307126,222856290,221092037,113051824],"submission":[2522,171037,39802792,142299,581,441,436,436,432,439],"total":2114140860,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":4,"nnode_src":0,"node":4},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[177070388,177104041,176963624,177257127,176602657,177342182,177529532,176907925,177274491,177624194],"completion":[177066979,176895902,176587759,176891954,176601153,177341713,177529115,176907537,177273357,177623774],"submission":[3409,208139,375865,365173,1504,469,417,388,1134,420],"total":1594620389,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":4,"nnode_src":0,"node":5},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[179519680,179297648,178765684,179513479,178535643,179579458,179894003,178869421,179550897,171430848],"completion":[179517362,179090369,178393187,179152450,178534978,179579024,179893537,178868997,179550483,171430388],"submission":[2318,207279,372497,361029,665,434,466,424,414,460],"total":1605446835,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":4,"nnode_src":0,"node":6},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[191864162,192140234,191583893,190886757,191706491,192357440,192228921,192223017,186225279,102364010],"completion":[191861987,191951931,191201374,190504198,191705881,192356990,192228479,192222583,186224878,102363217],"submission":[2175,188303,382519,382559,610,450,442,434,401,793],"total":1631729723,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":4,"nnode_src":0,"node":7},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[192771410,192227091,192000648,191850564,192564984,193139257,191919715,192277947,184529937,100985790],"completion":[192391200,191840250,191603221,191658889,192197396,193138703,191919289,192277520,184529525,100985280],"submission":[380210,386841,397427,191675,367588,554,426,427,412,510],"total":1631510865,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"hw"}
--- a/benchmarks/benchmark-results/copy-n0ton4-1gib-smart-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton4-1gib-smart-cpu-1e.json
@ -0,0 +1 @@
+{"count":2,"list":[{"affinity":{"nnode_dst":4,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[358682711,358651477,358654595,358632092,358683727,358646281,358751724,358190673,358735343,358849066],"completion":[104,104,103,104,103,104,121,105,106,105],"submission":[358682607,358651373,358654492,358631988,358683624,358646177,358751603,358190568,358735237,358848961],"total":3227816275,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":4,"nnode_src":0,"node":4},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[299559728,299432921,299492587,299506201,299538310,299452591,299555216,299504240,299526695,299629262],"completion":[120,99,114,110,102,106,115,114,107,104],"submission":[299559608,299432822,299492473,299506091,299538208,299452485,299555101,299504126,299526588,299629158],"total":2695650561,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"sw"}
--- a/benchmarks/benchmark-results/copy-n0ton5-1gib-allnodes-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton5-1gib-allnodes-cpu-1e.json
@ -0,0 +1 @@
+{"count":8,"list":[{"affinity":{"nnode_dst":5,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[364014626,236389437,215940138,217649994,216491640,217535995,218256137,217374980,215371650,204576964],"completion":[364011144,216376775,215818362,217532633,216491052,217535508,218255620,217374503,215371138,204576544],"submission":[3482,20012662,121776,117361,588,487,517,477,512,420],"total":1959601228,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":5,"nnode_src":0,"node":1},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[375925272,240697470,211684246,211530211,213388099,212501873,212423322,213374451,212707347,211809211],"completion":[375923039,210579260,211569667,211413495,213387466,212501431,212422886,213374059,212706903,211808788],"submission":[2233,30118210,114579,116716,633,442,436,392,444,423],"total":1940132220,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":5,"nnode_src":0,"node":2},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[316291290,229874272,226201922,225685846,225353164,225846663,225623072,226052741,225962125,164297557],"completion":[316140345,225040380,226084639,225685111,225352723,225846229,225622561,226052326,225961714,164297119],"submission":[150945,4833892,117283,735,441,434,511,415,411,438],"total":1974910547,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":5,"nnode_src":0,"node":3},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[329552844,229653317,223319245,223591024,223506662,222926569,222082235,222883543,223573964,180578175],"completion":[329550738,223127295,223204209,223468096,223506042,222926090,222081798,222883133,223573540,180577731],"submission":[2106,6526022,115036,122928,620,479,437,410,424,444],"total":1972128298,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":5,"nnode_src":0,"node":4},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[156491766,156131775,157168702,156938114,156963900,155987687,155994196,156835442,156382729,145711920],"completion":[156489117,155945819,156785935,156586885,156963339,155987219,155993587,156835047,156382321,145711490],"submission":[2649,185956,382767,351229,561,468,609,395,408,430],"total":1398125039,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":5,"nnode_src":0,"node":5},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[153726620,153254452,154034751,154264730,153888742,153005240,153259830,153273810,153205513,153080504],"completion":[153724106,153065285,153660687,153903730,153888231,153004816,153259155,153273423,153205124,153080063],"submission":[2514,189167,374064,361000,511,424,675,387,389,441],"total":1381281064,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":5,"nnode_src":0,"node":6},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[165260296,164967455,165325622,165693619,164711902,164579876,164320281,165399325,163713174,94816262],"completion":[164894908,164627781,164962432,165503563,164368375,164578761,164319865,165398886,163712750,94815699],"submission":[365388,339674,363190,190056,343527,1115,416,439,424,563],"total":1413542306,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":5,"nnode_src":0,"node":7},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[162631590,162949049,163018865,163939248,163156559,162866750,162134347,163176074,162585304,109487022],"completion":[162628990,162753740,162653929,163591052,163156027,162681388,162133857,163175661,162584903,109486589],"submission":[2600,195309,364936,348196,532,185362,490,413,401,433],"total":1413501923,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"hw"}
--- a/benchmarks/benchmark-results/copy-n0ton5-1gib-smart-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton5-1gib-smart-cpu-1e.json
@ -0,0 +1 @@
+{"count":2,"list":[{"affinity":{"nnode_dst":5,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[364856266,365091479,364810050,364895009,365131233,364762455,363783723,363227003,363143959,363410435],"completion":[137,98,101,100,101,102,171,100,100,168],"submission":[364856129,365091381,364809949,364894909,365131132,364762353,363783552,363226903,363143859,363410267],"total":3278302262,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":5,"nnode_src":0,"node":5},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[278292426,278212189,278267129,278182974,278204713,277430904,277449590,277404947,277276022,277342612],"completion":[109,104,128,102,102,100,102,100,292,101],"submission":[278292317,278212085,278267001,278182872,278204611,277430804,277449488,277404847,277275730,277342511],"total":2499782837,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"sw"}
--- a/benchmarks/benchmark-results/copy-n0ton6-1gib-allnodes-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton6-1gib-allnodes-cpu-1e.json
@ -0,0 +1 @@
+{"count":8,"list":[{"affinity":{"nnode_dst":6,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[307859780,226156202,223785840,224292158,224268815,223587944,223869572,224652677,224423193,181079892],"completion":[307857416,224188170,223669819,224173774,224268246,223587390,223869144,224652250,224422791,181079369],"submission":[2364,1968032,116021,118384,569,554,428,427,402,523],"total":1976130691,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":6,"nnode_src":0,"node":1},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[297712702,224973733,225024832,225741921,225025370,225364548,225148847,225190019,225273596,173049856],"completion":[297710608,224906441,224904467,225624280,225024831,225364125,225148431,225189609,225273191,173049418],"submission":[2094,67292,120365,117641,539,423,416,410,405,438],"total":1974805035,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":6,"nnode_src":0,"node":2},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[354385828,239442216,214437602,214413059,214708989,215179717,214446974,213909071,214110753,214346349],"completion":[354230004,213661621,214321501,214412458,214708499,215179293,214446544,213908032,214110351,214345890],"submission":[155824,25780595,116101,601,490,424,430,1039,402,459],"total":1955012605,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":6,"nnode_src":0,"node":3},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[342050094,233020979,216275807,216308781,216775401,216977880,217375055,216588513,216024362,210800292],"completion":[342047752,216428949,216157501,216188189,216774875,216977449,217374614,216588081,216023814,210799853],"submission":[2342,16592030,118306,120592,526,431,441,432,548,439],"total":1960160317,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":6,"nnode_src":0,"node":4},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[161861878,161609482,161907710,162028839,161833172,162187498,161847967,162239083,162554372,111151835],"completion":[161858516,161424932,161524671,161644104,161832668,162187047,161847520,162238683,162553967,111151418],"submission":[3362,184550,383039,384735,504,451,447,400,405,417],"total":1407372118,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":6,"nnode_src":0,"node":5},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[163766815,163528750,164067393,164052314,164319140,164104607,163548238,163757775,164203341,96838873],"completion":[163405454,163182811,163734556,163866315,163954205,164104106,163547801,163757374,164202944,96838431],"submission":[361361,345939,332837,185999,364935,501,437,401,397,442],"total":1408435149,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":6,"nnode_src":0,"node":6},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[152306273,152475054,153146960,152698806,152531882,153151120,152889531,152433220,152726964,152635311],"completion":[152302621,152289301,152770766,152171971,152531309,153150685,152889110,152432827,152726551,152634893],"submission":[3652,185753,376194,526835,573,435,421,393,413,418],"total":1375261654,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":6,"nnode_src":0,"node":7},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[155698733,155515631,156082316,155521655,155642174,156424218,155837110,155668177,155856883,145320356],"completion":[155696538,155311716,155733479,155163169,155641640,156423719,155836687,155667763,155856470,145319933],"submission":[2195,203915,348837,358486,534,499,423,414,413,423],"total":1391890393,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"hw"}
--- a/benchmarks/benchmark-results/copy-n0ton6-1gib-smart-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton6-1gib-smart-cpu-1e.json
@ -0,0 +1 @@
+{"count":2,"list":[{"affinity":{"nnode_dst":6,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[373537802,373414710,373556929,373551079,373407499,373468978,373659810,372494168,373396510,373509943],"completion":[44,114,105,104,103,105,100,102,102,104],"submission":[373537758,373414596,373556824,373550975,373407396,373468873,373659710,372494066,373396408,373509839],"total":3360473775,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":6,"nnode_src":0,"node":6},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[305928701,305856843,305872257,305763186,305814530,305884811,305799783,305846203,305934673,305945348],"completion":[113,102,111,104,126,105,102,135,105,220],"submission":[305928588,305856741,305872146,305763082,305814404,305884706,305799681,305846068,305934568,305945128],"total":2752736261,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"sw"}
--- a/benchmarks/benchmark-results/copy-n0ton7-1gib-allnodes-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton7-1gib-allnodes-cpu-1e.json
@ -0,0 +1 @@
+{"count":8,"list":[{"affinity":{"nnode_dst":7,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[327453386,236138453,225277435,224645474,225303768,226282145,225541407,225542884,225372922,162729055],"completion":[327450083,225288690,225163665,224532991,225303081,226281678,225540989,225542455,225372439,162728539],"submission":[3303,10849763,113770,112483,687,467,418,429,483,516],"total":1976847520,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":7,"nnode_src":0,"node":1},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[349489331,236928275,224388042,223285991,224090772,223389433,224237202,224602446,224228674,171565407],"completion":[349341480,223582680,224259990,223285444,224090320,223388995,224236766,224601969,224228269,171564930],"submission":[147851,13345595,128052,547,452,438,436,477,405,477],"total":1976729285,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":7,"nnode_src":0,"node":2},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[362941513,234465916,215759554,216193944,215191620,214828506,215209696,215155131,215233991,206929534],"completion":[362939238,214314442,215640099,216081561,215191048,214828057,215209267,215154675,215233563,206929036],"submission":[2275,20151474,119455,112383,572,449,429,456,428,498],"total":1948981563,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":7,"nnode_src":0,"node":3},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[376004681,241742898,211691796,211850225,211241573,212555074,212234041,211505281,211962119,211488856],"completion":[376002203,210304696,211571448,211732635,211241053,212554634,212233628,211504824,211961713,211488357],"submission":[2478,31438202,120348,117590,520,440,413,457,406,499],"total":1936285075,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":7,"nnode_src":0,"node":4},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[164265775,164242719,164489770,164357363,164792553,164333942,164258033,164889781,163802574,101211184],"completion":[163871503,163873395,164144859,164175524,164416819,164332681,164257536,164889372,163802179,101210741],"submission":[394272,369324,344911,181839,375734,1261,497,409,395,443],"total":1416390179,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":7,"nnode_src":0,"node":5},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[161281055,161552309,162582334,162645653,162753570,162569889,162105804,162587332,161913185,116333953],"completion":[161278881,161374041,162209110,162307446,162752993,162365279,162105254,162586924,161912786,116333507],"submission":[2174,178268,373224,338207,577,204610,550,408,399,446],"total":1415240333,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":7,"nnode_src":0,"node":6},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[156874003,156891269,157147112,157485116,157223042,157360913,157178540,157013509,157528483,145800599],"completion":[156871728,156412832,156767368,157138595,157222533,157360416,157177950,157013085,157528063,145800161],"submission":[2275,478437,379744,346521,509,497,590,424,420,438],"total":1403637975,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":7,"nnode_src":0,"node":7},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[153788366,153429039,154111456,154066657,154349816,154017879,154330138,153989814,154237171,153868114],"completion":[153786250,153242238,153743164,153710826,154349268,154017421,154329532,153989378,154236749,153867710],"submission":[2116,186801,368292,355831,548,458,606,436,422,404],"total":1386413009,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"hw"}
--- a/benchmarks/benchmark-results/copy-n0ton7-1gib-smart-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton7-1gib-smart-cpu-1e.json
@ -0,0 +1 @@
+{"count":2,"list":[{"affinity":{"nnode_dst":7,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[379377793,379336878,379284050,379340490,379331145,379472766,376914386,377610728,377760238,377738998],"completion":[105,182,105,102,102,121,104,105,103,104],"submission":[379377688,379336696,379283945,379340388,379331043,379472645,376914282,377610623,377760135,377738894],"total":3406814364,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":7,"nnode_src":0,"node":7},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[284330551,284349144,284154359,284314269,284176623,283659802,283505810,283609838,283562831,284326611],"completion":[106,258,110,180,102,102,101,100,133,103],"submission":[284330445,284348886,284154249,284314089,284176521,283659700,283505709,283609738,283562698,284326508],"total":2555676465,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"sw"}
--- a/benchmarks/benchmark-results/copy-n0ton8-1gib-allnodes-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton8-1gib-allnodes-cpu-1e.json
@ -0,0 +1 @@
+{"count":8,"list":[{"affinity":{"nnode_dst":8,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[98028701,98006190,98180929,98075298,98180565,98008728,97970868,97879080,98218848,98196709],"completion":[98026192,97886040,97934808,97842943,98179995,98008280,97970435,97878645,98218423,98196172],"submission":[2509,120150,246121,232355,570,448,433,435,425,537],"total":882837572,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":8,"nnode_src":0,"node":1},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[103077242,102832993,103077318,103154846,102789381,102937739,102905155,103166278,103320974,83074369],"completion":[102965823,102716332,102847727,103154308,102788982,102937282,102904711,103165874,103320579,83073472],"submission":[111419,116661,229591,538,399,457,444,404,395,897],"total":907608997,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":8,"nnode_src":0,"node":2},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[116641762,117007502,117059515,117203829,116899546,116903497,117275122,106634478,67991236,60838770],"completion":[116639715,116882239,116837078,116974461,116899027,116903067,117274716,106634056,67990419,60838257],"submission":[2047,125263,222437,229368,519,430,406,422,817,513],"total":937826794,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":8,"nnode_src":0,"node":3},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[116758722,117192519,117419350,117451489,116963587,117045121,117310185,105849114,63863863,57753769],"completion":[116756653,117064324,117184790,117234158,116963052,117044686,117309771,105848662,63863024,57753266],"submission":[2069,128195,234560,217331,535,435,414,452,839,503],"total":930861933,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":8,"nnode_src":0,"node":4},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[153008691,153132282,152958127,153189310,154463096,137217376,116656299,116623264,116612510,116632498],"completion":[152998117,153014435,152745764,152983580,154462601,137216562,116655710,116622832,116612112,116632085],"submission":[10574,117847,212363,205730,495,814,589,432,398,413],"total":1217579582,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":8,"nnode_src":0,"node":5},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[157657429,157595877,157574304,158024486,156393264,132264877,120017784,120045344,119957824,106089795],"completion":[157655738,157488308,157380074,157818924,156272507,132264260,120017317,120044945,119957422,106089355],"submission":[1691,107569,194230,205562,120757,617,467,399,402,440],"total":1228211368,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":8,"nnode_src":0,"node":6},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[169785224,169880242,169976186,171384833,152393730,133497725,133536806,133627231,106430299,68390006],"completion":[169782980,169769502,169770400,171174863,152392618,133497271,133536359,133626830,106429856,68389537],"submission":[2244,110740,205786,209970,1112,454,447,401,443,469],"total":1239217822,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":8,"nnode_src":0,"node":7},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[171303507,171016343,171103893,172691031,155711215,133448061,133535253,133413249,98929738,64656203],"completion":[171079959,170813739,170884403,172578949,150546928,133446940,133534790,133412791,98929289,64655609],"submission":[223548,202604,219490,112082,5164287,1121,463,458,449,594],"total":1234603871,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"hw"}
--- a/benchmarks/benchmark-results/copy-n0ton8-1gib-smart-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton8-1gib-smart-cpu-1e.json
@ -0,0 +1 @@
+{"count":4,"list":[{"affinity":{"nnode_dst":8,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[127771328,127728806,127690815,127716771,127711453,127731773,127719690,127863566,127758708,127753836],"completion":[122,148,103,119,107,105,294,106,107,111],"submission":[127771206,127728658,127690712,127716652,127711346,127731668,127719396,127863460,127758601,127753725],"total":1149688181,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":8,"nnode_src":0,"node":1},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[140296171,140345060,140342798,140284617,140333829,140323078,140291574,140290152,140108947,138939171],"completion":[104,103,101,100,100,161,101,160,127,98],"submission":[140296067,140344957,140342697,140284517,140333729,140322917,140291473,140289992,140108820,138939073],"total":1261271744,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":8,"nnode_src":0,"node":2},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[141072357,141105543,141100562,141046669,141027700,141086980,141090488,141049067,140830283,139862289],"completion":[101,99,17,101,99,252,253,102,105,104],"submission":[141072256,141105444,141100545,141046568,141027601,141086728,141090235,141048965,140830178,139862185],"total":1268213946,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":8,"nnode_src":0,"node":3},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[149442672,149359878,149346812,149471715,149403034,149448399,149363202,149454191,148473374,148383775],"completion":[104,102,103,103,101,103,102,103,103,103],"submission":[149442568,149359776,149346709,149471612,149402933,149448296,149363100,149454088,148473271,148383672],"total":1342716259,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"sw"}
--- a/benchmarks/benchmark-results/copy-n0ton9-1gib-allnodes-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton9-1gib-allnodes-cpu-1e.json
@ -0,0 +1 @@
+{"count":8,"list":[{"affinity":{"nnode_dst":9,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[96030175,95821940,95871725,96004080,96012384,95834653,95915617,95970244,95899390,95975559],"completion":[96027687,95739721,95691807,95838215,96011854,95834208,95915196,95969811,95898893,95975119],"submission":[2488,82219,179918,165865,530,445,421,433,497,440],"total":863316731,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":9,"nnode_src":0,"node":1},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[97104399,97185035,97240970,97433601,97253558,97203772,97301555,97197009,97270588,92996432],"completion":[97101988,97098152,97054936,97276092,97252962,97203328,97301147,97196610,97270191,92996015],"submission":[2411,86883,186034,157509,596,444,408,399,397,417],"total":871093649,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":9,"nnode_src":0,"node":2},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[98575667,98784671,98739470,98830724,98684362,98701818,98559748,98646275,98712937,83720969],"completion":[98573475,98668335,98548386,98669727,98683776,98701368,98559337,98645837,98712410,83720547],"submission":[2192,116336,191084,160997,586,450,411,438,527,422],"total":873390729,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":9,"nnode_src":0,"node":3},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[100078131,100068748,100020616,100046837,100150906,100072667,100220274,100104584,100054788,71787736],"completion":[100007743,99973466,99864785,100046267,100150455,100072215,100219840,100104124,100054379,71787308],"submission":[70388,95282,155831,570,451,452,434,460,409,428],"total":872537672,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":9,"nnode_src":0,"node":4},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[179827368,179996195,179965849,74180394,76898569,76943736,76921364,76932296,76992323,71704263],"completion":[179824882,179899951,179825072,73737968,76897218,76943185,76920940,76931875,76991918,71703845],"submission":[2486,96244,140777,442426,1351,551,424,421,405,418],"total":890819189,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":9,"nnode_src":0,"node":5},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[181368374,181593396,181886037,65188641,72768523,72814738,72845243,72817327,72772953,72810642],"completion":[181366268,181509066,181747543,57472145,72767229,72814173,72844816,72816932,72772527,72810230],"submission":[2106,84330,138494,7716496,1294,565,427,395,426,412],"total":865515447,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":9,"nnode_src":0,"node":6},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[183944790,183985726,181161227,112652274,86087232,86059117,85966766,86028575,62901594,42385198],"completion":[183800440,183853645,181019810,85968774,86009507,86058520,85966370,86028097,62901133,42384656],"submission":[144350,132081,141417,26683500,77725,597,396,478,461,542],"total":927304519,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":9,"nnode_src":0,"node":7},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[184377364,184215980,180040821,108747785,84038326,84048348,84109662,84070694,67287908,44037456],"completion":[184375271,184141626,179905034,84111268,84036952,84047858,84109235,84070295,67287502,44036980],"submission":[2093,74354,135787,24636517,1374,490,427,399,406,476],"total":920611228,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"hw"}
--- a/benchmarks/benchmark-results/copy-n0ton9-1gib-smart-cpu-1e.json
+++ b/benchmarks/benchmark-results/copy-n0ton9-1gib-smart-cpu-1e.json
@ -0,0 +1 @@
+{"count":4,"list":[{"affinity":{"nnode_dst":9,"nnode_src":0,"node":0},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[132387805,132405019,132384574,132353465,132370756,132339071,132336160,132374893,132331161,132384735],"completion":[106,138,123,106,100,102,101,101,110,104],"submission":[132387699,132404881,132384451,132353359,132370656,132338969,132336059,132374792,132331051,132384631],"total":1191292657,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":9,"nnode_src":0,"node":1},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[132531421,132538396,132593553,132496154,132228922,132069464,132100368,132107584,132077095,132071722],"completion":[103,140,99,104,141,108,103,102,105,104],"submission":[132531318,132538256,132593454,132496050,132228781,132069356,132100265,132107482,132076990,132071618],"total":1190295345,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":9,"nnode_src":0,"node":2},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[145982733,145902787,146038041,146017740,146000843,145988734,146035118,146036308,145384458,143710033],"completion":[55,103,100,105,102,104,102,292,103,104],"submission":[145982678,145902684,146037941,146017635,146000741,145988630,146035016,146036016,145384355,143709929],"total":1311127044,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}},{"affinity":{"nnode_dst":9,"nnode_src":0,"node":3},"report":{"iterations_completed":10,"status":"ok","time":{"combined":[146139229,146220329,146207059,146208700,146194644,146247060,146202126,146250784,145700971,144336207],"completion":[104,103,102,102,136,102,106,103,290,103],"submission":[146139125,146220226,146206957,146208598,146194508,146246958,146202020,146250681,145700681,144336104],"total":1313581548,"unit":"nanoseconds"}},"task":{"batching":{"batch_size":0,"batch_submit":false},"iterations":10,"size":1073741824}}],"path":"sw"}