kanshan/nccl-tests
1
0
Fork 0
mirror of https://github.com/NVIDIA/nccl-tests.git synced 2026-05-03 13:02:36 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

Modified warmup to run for more message sizes

Browse Source 
Loops between minBytes and maxBytes doubling size each time

Reduced default warmup iteration count to 1 (was 5)
This commit is contained in:
David Addison
2025-08-25 13:47:54 -07:00
parent fae7cb4727
commit f2015cbe82
2 changed files with 9 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files
README.md
+1 -1
View File
@@ -68,7 +68,7 @@ All tests support the same set of arguments :
* `-r,--root <root/all>` Specify which root to use. Only for operations with a root like broadcast or reduce. Default : 0. * `-r,--root <root/all>` Specify which root to use. Only for operations with a root like broadcast or reduce. Default : 0.
* Performance * Performance
* `-n,--iters <iteration count>` number of iterations. Default : 20. * `-n,--iters <iteration count>` number of iterations. Default : 20.
* `-w,--warmup_iters <warmup iteration count>` number of warmup iterations (not timed). Default : 5. * `-w,--warmup_iters <warmup iteration count>` number of warmup iterations (not timed). Default : 1.
* `-m,--agg_iters <aggregation count>` number of operations to aggregate together in each iteration. Default : 1. * `-m,--agg_iters <aggregation count>` number of operations to aggregate together in each iteration. Default : 1.
* `-N,--run_cycles <cycle count>` run & print each cycle. Default : 1; 0=infinite. * `-N,--run_cycles <cycle count>` run & print each cycle. Default : 1; 0=infinite.
* `-a,--average <0/1/2/3>` Report performance as an average across all ranks (MPI=1 only). <0=Rank0,1=Avg,2=Min,3=Max>. Default : 1. * `-a,--average <0/1/2/3>` Report performance as an average across all ranks (MPI=1 only). <0=Rank0,1=Avg,2=Min,3=Max>. Default : 1.
src/common.cu
+4 -9
View File
@@ -78,7 +78,7 @@ static size_t maxBytes = 32*1024*1024;
static size_t stepBytes = 1*1024*1024; static size_t stepBytes = 1*1024*1024;
static size_t stepFactor = 1; static size_t stepFactor = 1;
static int datacheck = 1; static int datacheck = 1;
static int warmup_iters = 5; static int warmup_iters = 1;
static int iters = 20; static int iters = 20;
static int agg_iters = 1; static int agg_iters = 1;
static int run_cycles = 1; static int run_cycles = 1;
@@ -607,19 +607,14 @@ testResult_t TimeTest(struct threadArgs* args, ncclDataType_t type, const char*
// Sync to avoid first-call timeout // Sync to avoid first-call timeout
Barrier(args); Barrier(args);
// Warm-up for large size // Warm-up for all sizes (using a stepfactor of 2)
setupArgs(args->maxbytes, type, args); for (size_t size = args->minbytes; size <= args->maxbytes; size = size * 2) {
setupArgs(size, type, args);
for (int iter = 0; iter < warmup_iters; iter++) { for (int iter = 0; iter < warmup_iters; iter++) {
TESTCHECK(startColl(args, type, op, root, 0, iter)); TESTCHECK(startColl(args, type, op, root, 0, iter));
} }
TESTCHECK(completeColl(args)); TESTCHECK(completeColl(args));
// Warm-up for small size
setupArgs(args->minbytes, type, args);
for (int iter = 0; iter < warmup_iters; iter++) {
TESTCHECK(startColl(args, type, op, root, 0, iter));
} }
TESTCHECK(completeColl(args));
// Benchmark // Benchmark
long repeat = run_cycles; long repeat = run_cycles;