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

WIP further MPI optimizations

Browse Source
This commit is contained in:
jpekkila
2020-03-24 19:02:58 +02:00
parent ef63813679
commit 672137f7f1
1 changed files with 177 additions and 131 deletions
Download Patch File Download Diff File Expand all files Collapse all files

56
src/core/device.cc
View File

@@ -1171,6 +1171,8 @@ acDeviceIntegrateMPI(const Device device, const AcReal dt)
timer_reset(&ttot);
MPI_Barrier(MPI_COMM_WORLD);
const int num_iterations = 1;
for (int i = 0; i < num_iterations; ++i) {
for (int isubstep = 0; isubstep < 3; ++isubstep) {
acPackCommData(device, corner_a0s, &corner_data);
acPackCommData(device, edgex_a0s, &edgex_data);
@@ -1267,6 +1269,7 @@ acDeviceIntegrateMPI(const Device device, const AcReal dt)
acDeviceSynchronizeStream(device, STREAM_ALL); // Wait until inner and outer done
////////////////////////////////////////////
}
}
cudaDeviceSynchronize();
MPI_Barrier(MPI_COMM_WORLD);
@@ -1277,7 +1280,8 @@ acDeviceIntegrateMPI(const Device device, const AcReal dt)
MPI_Comm_rank(MPI_COMM_WORLD, &pid);
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
if (!pid) {
printf("--- Total communication time per step: %f ms\n", msec);
printf("--- Total communication time per step w/ integration: %f ms\n",
msec / num_iterations);
// Write out to file
FILE* fp = fopen("benchmark.result", "a+");
@@ -1434,7 +1438,6 @@ acDeviceCommunicateHalosMPI(const Device device)
timer_reset(&ttot);
MPI_Barrier(MPI_COMM_WORLD);
for (int isubstep = 0; isubstep < 3; ++isubstep) {
acPackCommData(device, corner_a0s, &corner_data);
acPackCommData(device, edgex_a0s, &edgex_data);
acPackCommData(device, edgey_a0s, &edgey_data);
@@ -1486,7 +1489,6 @@ acDeviceCommunicateHalosMPI(const Device device)
acUnpackCommData(device, sidexy_b0s, &sidexy_data);
acUnpackCommData(device, sidexz_b0s, &sidexz_data);
acUnpackCommData(device, sideyz_b0s, &sideyz_data);
}
cudaDeviceSynchronize();
MPI_Barrier(MPI_COMM_WORLD);
@@ -1497,7 +1499,7 @@ acDeviceCommunicateHalosMPI(const Device device)
MPI_Comm_rank(MPI_COMM_WORLD, &pid);
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
if (!pid) {
printf("--- Total communication time per step: %f ms\n", msec);
printf("--- Total communication time per substep (comm): %f ms\n", msec);
// Write out to file
FILE* fp = fopen("benchmark.result", "a+");
@@ -1517,6 +1519,13 @@ acDeviceCommunicateHalosMPI(const Device device)
return AC_SUCCESS;
}
/*
static int3
findOptimalDecomposition(const int3 nn)
{
int3 decomposition = (int3){1, 1, 1};
}*/
AcResult
acDeviceRunMPITest(void)
{
@@ -1632,7 +1641,6 @@ acDeviceRunMPITest(void)
// VERIFY ////////////////////////////////////////////////////
if (pid == 0) {
// acMeshApplyPeriodicBounds(&model);
acModelIntegrateStep(model, FLT_EPSILON);
acMeshApplyPeriodicBounds(&model);
@@ -1657,3 +1665,41 @@ acDeviceRunMPITest(void)
return AC_FAILURE;
}
#endif // AC_MPI_ENABLED
/*
struct grid_s {
Device device;
};
typedef grid_s* Grid;
AcResult
acGridInit(void)
{
MPI_Init(NULL, NULL);
int nprocs, pid;
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
MPI_Comm_rank(MPI_COMM_WORLD, &pid);
char processor_name[MPI_MAX_PROCESSOR_NAME];
int name_len;
MPI_Get_processor_name(processor_name, &name_len);
printf("Processor %s. Process %d of %d.\n", processor_name, pid, nprocs);
}
AcResult
acGridLoad(const AcMesh mesh, Grid* grid)
{
}
AcResult
acGridStore(const Grid grid, AcMesh* mesh)
{
}
AcResult
acGridQuit(AcGrid& grid)
{
}
*/