cwpearson/astaroth
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Added timing to the MPI benchmark

Browse Source
This commit is contained in:
jpekkila
2019-10-17 17:43:54 +03:00
parent e0a631d81a
commit 3d852e5082
1 changed files with 32 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

46
src/core/device.cu
View File

@@ -975,6 +975,7 @@ acHostCommunicateHalosMPI(AcMesh* submesh)
// From Astaroth Utils // From Astaroth Utils
#include "src/utils/config_loader.h" #include "src/utils/config_loader.h"
#include "src/utils/memory.h" #include "src/utils/memory.h"
#include "src/utils/timer_hires.h"
#include "src/utils/verification.h" #include "src/utils/verification.h"
// --smpiargs="-gpu" // --smpiargs="-gpu"
AcResult AcResult
@@ -1017,6 +1018,10 @@ acDeviceRunMPITest(void)
AcMeshInfo info; AcMeshInfo info;
acLoadConfig(AC_DEFAULT_CONFIG, &info); acLoadConfig(AC_DEFAULT_CONFIG, &info);
const int nn = 256;
info.int_params[AC_nx] = info.int_params[AC_ny] = info.int_params[AC_nz] = nn;
acUpdateConfig(&info);
AcMesh model, candidate; AcMesh model, candidate;
// Master CPU // Master CPU
@@ -1060,23 +1065,36 @@ acDeviceRunMPITest(void)
acDeviceCreate(0, submesh_info, &device); acDeviceCreate(0, submesh_info, &device);
acDeviceLoadMesh(device, STREAM_DEFAULT, submesh); acDeviceLoadMesh(device, STREAM_DEFAULT, submesh);
///// Communication start ////////////////////////////// Timer start
{ const int num_iters = 100;
const int3 start = (int3){0, 0, NGHOST}; Timer total_time;
const int3 end = (int3){subgrid_m.x, subgrid_m.y, subgrid_m.z - NGHOST}; timer_reset(&total_time);
acDevicePeriodicBoundconds(device, STREAM_DEFAULT, start, end); for (int i = 0; i < num_iters; ++i) {
} ///// Communication start
{
const int3 start = (int3){0, 0, NGHOST};
const int3 end = (int3){subgrid_m.x, subgrid_m.y, subgrid_m.z - NGHOST};
acDevicePeriodicBoundconds(device, STREAM_DEFAULT, start, end);
}
#if 1 // GPU-GPU if CUDA-aware MPI, otherwise managed CPU-GPU-GPU-CPU #if 1 // GPU-GPU if CUDA-aware MPI, otherwise managed CPU-GPU-GPU-CPU
acDeviceSynchronizeStream(device, STREAM_DEFAULT); acDeviceSynchronizeStream(device, STREAM_DEFAULT);
MPI_Barrier(MPI_COMM_WORLD); MPI_Barrier(MPI_COMM_WORLD);
acDeviceCommunicateHalosMPI(device); // Includes periodic bounds at first and last ghost zone acDeviceCommunicateHalosMPI(
MPI_Barrier(MPI_COMM_WORLD); device); // Includes periodic bounds at first and last ghost zone
MPI_Barrier(MPI_COMM_WORLD);
#else // Explicit GPU-CPU-CPU-GPU #else // Explicit GPU-CPU-CPU-GPU
acDeviceStoreMesh(device, STREAM_DEFAULT, &submesh); acDeviceStoreMesh(device, STREAM_DEFAULT, &submesh);
acHostCommunicateHalosMPI(&submesh); acHostCommunicateHalosMPI(&submesh);
acDeviceLoadMesh(device, STREAM_DEFAULT, submesh); acDeviceLoadMesh(device, STREAM_DEFAULT, submesh);
#endif #endif
///// Communication end ///// Communication end
}
if (pid == 0) {
const double ms_elapsed = timer_diff_nsec(total_time) / 1e6;
printf("vertices: %d^3, iterations: %d\n", nn, num_iters);
printf("Total time: %f ms\n", ms_elapsed);
}
////////////////////////////// Timer end
acDeviceStoreMesh(device, STREAM_DEFAULT, &submesh); acDeviceStoreMesh(device, STREAM_DEFAULT, &submesh);
acDeviceDestroy(device); acDeviceDestroy(device);