make load balancing thread-local to reduce atomic cost

qdp_project/src/utils/BenchmarkHelpers.cpp

@@ -38,13 +38,13 @@ std::vector<int> CopyMethodPolicy(const int numa_dst_node, const int numa_src_no
     // as submission cost matters for low transfer size
 
     if (data_size < 16_MiB) {
-        static std::atomic<int> last_node = 0;
-        const int node = last_node.fetch_add(1) % 4;
+        static thread_local int last_node = 0;
+        const int node = last_node++ % 4;
         return std::vector<int>{ node };
     }
     else {
-        static std::atomic<int> last_group = 0;
-        const int group = last_group.fetch_add(1) % 2;
+        static thread_local int last_group = 0;
+        const int group = last_group++ % 2;
         return group == 0 ? std::vector<int>{ 0, 1 } : std::vector<int>{ 2, 3 };
     }
 }