move mode selection to Configuration.hpp, adapt the CopyMethodPolicy-Function to return only src_node for task sizes under 16MiB which is now required to not cause high submission count which slows down small copies

qdp_project/src/Benchmark.cpp

@@ -14,10 +14,6 @@
 
 #include "../../offloading-cacher/cache.hpp"
 
-#ifndef MODE_SET_BY_CMAKE
-#define MODE_SIMPLE_PREFETCH
-#endif
-
 #include "Configuration.hpp"
 #include "BenchmarkHelpers.cpp"
 

qdp_project/src/Configuration.hpp

@@ -1,5 +1,11 @@
 #pragma once
 
+#include "utils/memory_literals.h"
+
+#ifndef MODE_SET_BY_CMAKE
+#define MODE_SIMPLE_PREFETCH
+#endif
+
 constexpr size_t WL_SIZE_B = 4_GiB;
 constexpr uint32_t WARMUP_ITERATION_COUNT = 5;
 constexpr uint32_t ITERATION_COUNT = 5;

qdp_project/src/utils/BenchmarkHelpers.cpp

@@ -2,6 +2,8 @@
 
 #include "../Configuration.hpp"
 
+#include "memory_literals.h"
+
 constexpr size_t SCANA_TIMING_INDEX = 0;
 constexpr size_t SCANB_TIMING_INDEX = 1;
 constexpr size_t AGGRJ_TIMING_INDEX = 2;
@@ -38,6 +40,14 @@ int CachePlacementPolicy(const int numa_dst_node, const int numa_src_node, const
 }
 
 std::vector<int> CopyMethodPolicy(const int numa_dst_node, const int numa_src_node, const size_t data_size) {
+    // for small data it is more efficient to run on only one node
+    // which causes less submissions and therefore completes faster
+    // as submission cost matters for low transfer size
+
+    if (data_size < 16_MiB) {
+        return std::vector<int>{ (numa_src_node >= 8 ? numa_src_node - 8 : numa_src_node) };
+    }
+
     // for sufficiently large data, smart copy is used which will utilize
     // all four engines for intra-socket copy operations and cross copy on
     // the source and destination nodes for inter-socket copy