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

Pulled useful changes from the benchmark branch. GPUDirect RDMA (unpinned) is now the default for MPI communication.

Browse Source
This commit is contained in:
jpekkila
2020-07-29 16:39:24 +03:00
parent 6cab3586cf
commit 003c202e8c
4 changed files with 238 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
samples/benchmark/main.cc
View File

@@ -207,24 +207,18 @@ main(int argc, char** argv)
results[nth_percentile * num_iters], 100 * nth_percentile);
char path[4096] = "";
if (test == TEST_STRONG_SCALING)
strncpy(path, "strong_scaling.csv", sizeof(path));
else if (test == TEST_WEAK_SCALING)
strncpy(path, "weak_scaling.csv", sizeof(path));
else
ERROR("Invalid test type");
sprintf(path, "%s_%d.csv", test == TEST_STRONG_SCALING ? "strong" : "weak", nprocs);
FILE* fp = fopen(path, "a");
ERRCHK_ALWAYS(fp);
// Format
// nprocs, measured (ms)
fprintf(fp, "%d, %g\n", nprocs, results[nth_percentile * num_iters]);
// nprocs, min, 50th perc, 90th perc, max
fprintf(fp, "%d, %g, %g, %g, %g\n", nprocs, results[0], results[0.5 * num_iters], results[nth_percentile * num_iters], results[num_iters-1]);
fclose(fp);
}
/*
const size_t num_iters = 100;
const size_t num_iters = 1000;
const double nth_percentile = 0.90;
std::vector<double> results; // ms

15
samples/genbenchmarkscripts/main.c
View File

@@ -21,6 +21,7 @@ main(void)
fprintf(fp, "#SBATCH --time=00:14:59\n");
fprintf(fp, "#SBATCH --mem=32000\n");
fprintf(fp, "#SBATCH --partition=gpu\n");
fprintf(fp, "#SBATCH --cpus-per-task=10\n");
// nprocs, nodes, gpus
const int max_gpus_per_node = 4;
@@ -29,22 +30,30 @@ main(void)
fprintf(fp, "#SBATCH --gres=gpu:v100:%d\n", gpus_per_node);
fprintf(fp, "#SBATCH -n %d\n", nprocs);
fprintf(fp, "#SBATCH -N %d\n", nodes);
fprintf(fp, "#SBATCH --exclusive\n");
//fprintf(fp, "#SBATCH --exclusive\n");
if (nprocs > 4)
fprintf(fp, "#SBATCH --ntasks-per-socket=2\n");
// Modules
fprintf(fp, "module load gcc/8.3.0 cuda/10.1.168 cmake hpcx-mpi/2.5.0-cuda nccl\n");
// OpenMPI
fprintf(fp, "module load gcc/8.3.0 cuda/10.1.168 cmake openmpi nccl\n");
// HPCX
//fprintf(fp, "module load gcc/8.3.0 cuda/10.1.168 cmake hpcx-mpi/2.5.0-cuda nccl\n");
fprintf(fp, "export UCX_MEMTYPE_CACHE=n\n");
// Profile and run
fprintf(fp, "mkdir -p profile_%d\n", nprocs);
//fprintf(fp, "mkdir -p profile_%d\n", nprocs);
const int nx = 256; // max size 1792;
const int ny = nx;
const int nz = nx;
/*
fprintf(fp,
//"srun nvprof --annotate-mpi openmpi -o profile_%d/%%p.nvprof ./benchmark %d %d "
//"%d\n",
"srun ./benchmark %d %d %d\n", nx, ny, nz);
*/
fprintf(fp, "srun ./benchmark %d %d %d\n", nx, ny, nz);
fclose(fp);
}