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

Adjust burn GPU ramp timing by profile

Browse Source
This commit is contained in:
Michael Chus
2026-04-12 15:58:30 +03:00
parent 525ed8b8fc
commit e16d0f34b5
3 changed files with 156 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files

90
audit/internal/webui/tasks.go
View File

@@ -152,6 +152,12 @@ type burnPreset struct {
DurationSec int
}
type nvidiaRampSpec struct {
DurationSec int
StaggerSeconds int
TotalDurationSec int
}
func resolveBurnPreset(profile string) burnPreset {
switch profile {
case "overnight":
@@ -163,11 +169,43 @@ func resolveBurnPreset(profile string) burnPreset {
}
}
func boolToNvidiaStaggerSeconds(enabled bool, selected []int) int {
if enabled && len(selected) > 1 {
return 180
func resolveNvidiaRampPlan(profile string, enabled bool, selected []int) (nvidiaRampSpec, error) {
base := resolveBurnPreset(profile).DurationSec
plan := nvidiaRampSpec{
DurationSec: base,
TotalDurationSec: base,
}
return 0
if !enabled {
return plan, nil
}
count := len(selected)
if count == 0 {
return nvidiaRampSpec{}, fmt.Errorf("staggered NVIDIA burn requires explicit GPU selection")
}
if count == 1 {
return plan, nil
}
switch profile {
case "acceptance":
plan.StaggerSeconds = 10 * 60
plan.TotalDurationSec = plan.DurationSec + plan.StaggerSeconds*(count-1)
case "overnight":
plan.StaggerSeconds = 60 * 60
plan.TotalDurationSec = 8 * 60 * 60
minTotal := count * 60 * 60
if plan.TotalDurationSec < minTotal {
plan.TotalDurationSec = minTotal
}
if plan.TotalDurationSec > 10*60*60 {
return nvidiaRampSpec{}, fmt.Errorf("overnight staggered NVIDIA burn supports at most 10 GPUs")
}
plan.DurationSec = plan.TotalDurationSec - plan.StaggerSeconds*(count-1)
default:
plan.StaggerSeconds = 2 * 60
plan.TotalDurationSec = plan.DurationSec + plan.StaggerSeconds*(count-1)
}
return plan, nil
}
func resolvePlatformStressPreset(profile string) platform.PlatformStressOptions {
@@ -600,7 +638,7 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
RunNCCL: t.params.RunNCCL,
ParallelGPUs: t.params.ParallelGPUs,
}, j.append)
case "nvidia-compute":
case "nvidia-compute":
if a == nil {
err = fmt.Errorf("app not configured")
break
@@ -609,11 +647,18 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
if t.params.BurnProfile != "" && dur <= 0 {
dur = resolveBurnPreset(t.params.BurnProfile).DurationSec
}
staggerSec := boolToNvidiaStaggerSeconds(t.params.StaggerGPUStart, t.params.GPUIndices)
if staggerSec > 0 {
j.append(fmt.Sprintf("NVIDIA staggered ramp-up enabled: %ds per GPU", staggerSec))
}
archive, err = a.RunNvidiaOfficialComputePack(ctx, "", dur, t.params.GPUIndices, staggerSec, j.append)
rampPlan, planErr := resolveNvidiaRampPlan(t.params.BurnProfile, t.params.StaggerGPUStart, t.params.GPUIndices)
if planErr != nil {
err = planErr
break
}
if t.params.BurnProfile != "" && t.params.StaggerGPUStart && dur <= 0 {
dur = rampPlan.DurationSec
}
if rampPlan.StaggerSeconds > 0 {
j.append(fmt.Sprintf("NVIDIA staggered ramp-up enabled: %ds per GPU; post-ramp hold: %ds; total runtime: %ds", rampPlan.StaggerSeconds, dur, rampPlan.TotalDurationSec))
}
archive, err = a.RunNvidiaOfficialComputePack(ctx, "", dur, t.params.GPUIndices, rampPlan.StaggerSeconds, j.append)
case "nvidia-targeted-power":
if a == nil {
err = fmt.Errorf("app not configured")
@@ -663,13 +708,24 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
if t.params.BurnProfile != "" && dur <= 0 {
dur = resolveBurnPreset(t.params.BurnProfile).DurationSec
}
archive, err = runNvidiaStressPackCtx(a, ctx, "", platform.NvidiaStressOptions{
DurationSec: dur,
Loader: t.params.Loader,
GPUIndices: t.params.GPUIndices,
ExcludeGPUIndices: t.params.ExcludeGPUIndices,
StaggerSeconds: boolToNvidiaStaggerSeconds(t.params.StaggerGPUStart, t.params.GPUIndices),
}, j.append)
rampPlan, planErr := resolveNvidiaRampPlan(t.params.BurnProfile, t.params.StaggerGPUStart, t.params.GPUIndices)
if planErr != nil {
err = planErr
break
}
if t.params.BurnProfile != "" && t.params.StaggerGPUStart && dur <= 0 {
dur = rampPlan.DurationSec
}
if rampPlan.StaggerSeconds > 0 {
j.append(fmt.Sprintf("NVIDIA staggered ramp-up enabled: %ds per GPU; post-ramp hold: %ds; total runtime: %ds", rampPlan.StaggerSeconds, dur, rampPlan.TotalDurationSec))
}
archive, err = runNvidiaStressPackCtx(a, ctx, "", platform.NvidiaStressOptions{
DurationSec: dur,
Loader: t.params.Loader,
GPUIndices: t.params.GPUIndices,
ExcludeGPUIndices: t.params.ExcludeGPUIndices,
StaggerSeconds: rampPlan.StaggerSeconds,
}, j.append)
case "memory":
if a == nil {
err = fmt.Errorf("app not configured")