reanimator/bee
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 1 Wiki Activity

fix(stress): keep platform burn responsive under load

Browse Source
This commit is contained in:
Michael Chus
2026-03-31 22:28:26 +03:00
parent ea660500c9
commit c9ee078622
3 changed files with 141 additions and 41 deletions
Download Patch File Download Diff File Expand all files Collapse all files

91
iso/builder/bee-gpu-stress.c
View File

@@ -36,6 +36,7 @@ typedef void *CUstream;
#define MAX_CUBLAS_PROFILES 5
#define MIN_PROFILE_BUDGET_BYTES ((size_t)4u * 1024u * 1024u)
#define MIN_STREAM_BUDGET_BYTES ((size_t)64u * 1024u * 1024u)
#define STRESS_LAUNCH_DEPTH 8
static const char *ptx_source =
".version 6.0\n"
@@ -422,24 +423,31 @@ static int run_ptx_fallback(struct cuda_api *api,
double deadline = start + (double)seconds;
while (now_seconds() < deadline) {
launches_per_wave = 0;
for (int lane = 0; lane < stream_count; lane++) {
unsigned int blocks = (unsigned int)((words[lane] + threads - 1) / threads);
if (!check_rc(api,
"cuLaunchKernel",
api->cuLaunchKernel(kernel,
blocks,
1,
1,
threads,
1,
1,
0,
streams[lane],
params[lane],
NULL))) {
goto fail;
for (int depth = 0; depth < STRESS_LAUNCH_DEPTH && now_seconds() < deadline; depth++) {
int launched_this_batch = 0;
for (int lane = 0; lane < stream_count; lane++) {
unsigned int blocks = (unsigned int)((words[lane] + threads - 1) / threads);
if (!check_rc(api,
"cuLaunchKernel",
api->cuLaunchKernel(kernel,
blocks,
1,
1,
threads,
1,
1,
0,
streams[lane],
params[lane],
NULL))) {
goto fail;
}
launches_per_wave++;
launched_this_batch++;
}
if (launched_this_batch <= 0) {
break;
}
launches_per_wave++;
}
if (launches_per_wave <= 0) {
goto fail;
@@ -460,10 +468,11 @@ static int run_ptx_fallback(struct cuda_api *api,
report->iterations = iterations;
snprintf(report->details,
sizeof(report->details),
"fallback_int32=OK requested_mb=%d actual_mb=%d streams=%d per_stream_mb=%zu iterations=%lu\n",
"fallback_int32=OK requested_mb=%d actual_mb=%d streams=%d queue_depth=%d per_stream_mb=%zu iterations=%lu\n",
size_mb,
report->buffer_mb,
report->stream_count,
STRESS_LAUNCH_DEPTH,
bytes_per_stream[0] / (1024u * 1024u),
iterations);
@@ -1184,10 +1193,11 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
report->buffer_mb = (int)(total_budget / (1024u * 1024u));
append_detail(report->details,
sizeof(report->details),
"requested_mb=%d actual_mb=%d streams=%d mp_count=%d per_worker_mb=%zu\n",
"requested_mb=%d actual_mb=%d streams=%d queue_depth=%d mp_count=%d per_worker_mb=%zu\n",
size_mb,
report->buffer_mb,
report->stream_count,
STRESS_LAUNCH_DEPTH,
mp_count,
per_profile_budget / (1024u * 1024u));
@@ -1239,26 +1249,33 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
double deadline = now_seconds() + (double)seconds;
while (now_seconds() < deadline) {
wave_launches = 0;
for (int i = 0; i < prepared_count; i++) {
if (!prepared[i].ready) {
continue;
}
if (!run_cublas_profile(handle, &cublas, &prepared[i])) {
append_detail(report->details,
sizeof(report->details),
"%s=FAILED runtime\n",
prepared[i].desc.name);
for (int j = 0; j < prepared_count; j++) {
destroy_profile(&cublas, cuda, &prepared[j]);
for (int depth = 0; depth < STRESS_LAUNCH_DEPTH && now_seconds() < deadline; depth++) {
int launched_this_batch = 0;
for (int i = 0; i < prepared_count; i++) {
if (!prepared[i].ready) {
continue;
}
cublas.cublasLtDestroy(handle);
destroy_streams(cuda, streams, stream_count);
cuda->cuCtxDestroy(ctx);
return 0;
if (!run_cublas_profile(handle, &cublas, &prepared[i])) {
append_detail(report->details,
sizeof(report->details),
"%s=FAILED runtime\n",
prepared[i].desc.name);
for (int j = 0; j < prepared_count; j++) {
destroy_profile(&cublas, cuda, &prepared[j]);
}
cublas.cublasLtDestroy(handle);
destroy_streams(cuda, streams, stream_count);
cuda->cuCtxDestroy(ctx);
return 0;
}
prepared[i].iterations++;
report->iterations++;
wave_launches++;
launched_this_batch++;
}
if (launched_this_batch <= 0) {
break;
}
prepared[i].iterations++;
report->iterations++;
wave_launches++;
}
if (wave_launches <= 0) {
break;