Grid/Benchmark_Grid.cpp

@@ -264,10 +264,9 @@ class Benchmark
     return;
   }
 
-  static void Latency(void)
+  static void PointToPoint(void)
   {
-    int Nwarmup = 100;
-    int Nloop = 1000;
+    int Nloop = 200;
 
     Coordinate simd_layout = GridDefaultSimd(Nd, vComplexD::Nsimd());
     Coordinate mpi_layout = GridDefaultMpi();
@@ -292,7 +291,7 @@ class Benchmark
     int bytes = 8;
     void *buf_from = acceleratorAllocDevice(bytes);
     void *buf_to = acceleratorAllocDevice(bytes);
-    nlohmann::json json_latency;
+    nlohmann::json json_p2p;
     for (int from = 0; from < ranks; ++from)
       for (int to = 0; to < ranks; ++to)
       {
@@ -303,7 +302,7 @@ class Benchmark
         time_statistics timestat;
         MPI_Status status;
 
-        for (int i = -Nwarmup; i < Nloop; ++i)
+        for (int i = 0; i < Nloop; ++i)
         {
           double start = usecond();
           if (from == me)
@@ -322,7 +321,6 @@ class Benchmark
             assert(err == MPI_SUCCESS);
           }
           double stop = usecond();
-          if (i >= 0)
           t_time[i] = stop - start;
         }
         // important: only the 'from' rank has a trustworthy time
@@ -337,9 +335,9 @@ class Benchmark
         tmp["time_usec"] = timestat.mean;
         tmp["time_usec_error"] = timestat.err;
         tmp["time_usec_max"] = timestat.min;
-        json_latency.push_back(tmp);
+        json_p2p.push_back(tmp);
       }
-    json_results["latency"] = json_latency;
+    json_results["latency"] = json_p2p;
 
     acceleratorFreeDevice(buf_from);
     acceleratorFreeDevice(buf_to);
@@ -886,7 +884,7 @@ int main(int argc, char **argv)
   int do_su4 = 1;
   int do_memory = 1;
   int do_comms = 1;
-  int do_latency = 1;
+  int do_p2p = 1;
   int do_flops = 1;
   int Ls = 1;
 
@@ -922,12 +920,12 @@ int main(int argc, char **argv)
     Benchmark::Comms();
   }
 
-  if (do_latency)
+  if (do_p2p)
   {
     grid_big_sep();
-    std::cout << GridLogMessage << " Latency benchmark " << std::endl;
+    std::cout << GridLogMessage << " Point-to-Point benchmark " << std::endl;
     grid_big_sep();
-    Benchmark::Latency();
+    Benchmark::PointToPoint();
   }
 
   if (do_flops)