Split bee-bench into perf and power workflows
This commit is contained in:
Mikhail Chusavitin
@@ -110,7 +110,7 @@ func writeTaskRunResponse(w http.ResponseWriter, tasks []*Task) {
|
||||
|
||||
func shouldSplitHomogeneousNvidiaTarget(target string) bool {
|
||||
switch strings.TrimSpace(target) {
|
||||
case "nvidia", "nvidia-targeted-stress", "nvidia-benchmark", "nvidia-compute",
|
||||
case "nvidia", "nvidia-targeted-stress", "nvidia-bench-perf", "nvidia-bench-power", "nvidia-compute",
|
||||
"nvidia-targeted-power", "nvidia-pulse", "nvidia-interconnect",
|
||||
"nvidia-bandwidth", "nvidia-stress":
|
||||
return true
|
||||
@@ -127,7 +127,7 @@ func defaultTaskPriority(target string, params taskParams) int {
|
||||
return taskPriorityInstallToRAM
|
||||
case "audit":
|
||||
return taskPriorityAudit
|
||||
case "nvidia-benchmark":
|
||||
case "nvidia-bench-perf", "nvidia-bench-power":
|
||||
return taskPriorityBenchmark
|
||||
case "nvidia-stress", "amd-stress", "memory-stress", "sat-stress", "platform-stress", "nvidia-compute":
|
||||
return taskPriorityBurn
|
||||
@@ -573,131 +573,142 @@ func (h *handler) handleAPISATRun(target string) http.HandlerFunc {
|
||||
}
|
||||
}
|
||||
|
||||
func (h *handler) handleAPIBenchmarkNvidiaRun(w http.ResponseWriter, r *http.Request) {
|
||||
if h.opts.App == nil {
|
||||
writeError(w, http.StatusServiceUnavailable, "app not configured")
|
||||
return
|
||||
}
|
||||
|
||||
var body struct {
|
||||
Profile string `json:"profile"`
|
||||
SizeMB int `json:"size_mb"`
|
||||
GPUIndices []int `json:"gpu_indices"`
|
||||
ExcludeGPUIndices []int `json:"exclude_gpu_indices"`
|
||||
RunNCCL *bool `json:"run_nccl"`
|
||||
ParallelGPUs *bool `json:"parallel_gpus"`
|
||||
RampUp *bool `json:"ramp_up"`
|
||||
DisplayName string `json:"display_name"`
|
||||
}
|
||||
if r.Body != nil {
|
||||
if err := json.NewDecoder(r.Body).Decode(&body); err != nil && !errors.Is(err, io.EOF) {
|
||||
writeError(w, http.StatusBadRequest, "invalid request body")
|
||||
func (h *handler) handleAPIBenchmarkNvidiaRunKind(target string) http.HandlerFunc {
|
||||
return func(w http.ResponseWriter, r *http.Request) {
|
||||
if h.opts.App == nil {
|
||||
writeError(w, http.StatusServiceUnavailable, "app not configured")
|
||||
return
|
||||
}
|
||||
}
|
||||
|
||||
runNCCL := true
|
||||
if body.RunNCCL != nil {
|
||||
runNCCL = *body.RunNCCL
|
||||
}
|
||||
parallelGPUs := false
|
||||
if body.ParallelGPUs != nil {
|
||||
parallelGPUs = *body.ParallelGPUs
|
||||
}
|
||||
rampUp := false
|
||||
if body.RampUp != nil {
|
||||
rampUp = *body.RampUp
|
||||
}
|
||||
// Build a descriptive base name that includes profile and mode so the task
|
||||
// list is self-explanatory without opening individual task detail pages.
|
||||
profile := strings.TrimSpace(body.Profile)
|
||||
if profile == "" {
|
||||
profile = "standard"
|
||||
}
|
||||
name := taskDisplayName("nvidia-benchmark", "", "")
|
||||
if strings.TrimSpace(body.DisplayName) != "" {
|
||||
name = body.DisplayName
|
||||
}
|
||||
// Append profile tag.
|
||||
name = fmt.Sprintf("%s · %s", name, profile)
|
||||
var body struct {
|
||||
Profile string `json:"profile"`
|
||||
SizeMB int `json:"size_mb"`
|
||||
GPUIndices []int `json:"gpu_indices"`
|
||||
ExcludeGPUIndices []int `json:"exclude_gpu_indices"`
|
||||
RunNCCL *bool `json:"run_nccl"`
|
||||
ParallelGPUs *bool `json:"parallel_gpus"`
|
||||
RampUp *bool `json:"ramp_up"`
|
||||
DisplayName string `json:"display_name"`
|
||||
}
|
||||
if r.Body != nil {
|
||||
if err := json.NewDecoder(r.Body).Decode(&body); err != nil && !errors.Is(err, io.EOF) {
|
||||
writeError(w, http.StatusBadRequest, "invalid request body")
|
||||
return
|
||||
}
|
||||
}
|
||||
|
||||
if rampUp && len(body.GPUIndices) > 1 {
|
||||
// Ramp-up mode: resolve GPU list, then create one task per prefix
|
||||
// [gpu0], [gpu0,gpu1], ..., [gpu0,...,gpuN-1], each running in parallel.
|
||||
gpus, err := apiListNvidiaGPUs(h.opts.App)
|
||||
if err != nil {
|
||||
writeError(w, http.StatusBadRequest, err.Error())
|
||||
runNCCL := true
|
||||
if body.RunNCCL != nil {
|
||||
runNCCL = *body.RunNCCL
|
||||
}
|
||||
parallelGPUs := false
|
||||
if body.ParallelGPUs != nil {
|
||||
parallelGPUs = *body.ParallelGPUs
|
||||
}
|
||||
rampUp := false
|
||||
if body.RampUp != nil {
|
||||
rampUp = *body.RampUp
|
||||
}
|
||||
// Build a descriptive base name that includes profile and mode so the task
|
||||
// list is self-explanatory without opening individual task detail pages.
|
||||
profile := strings.TrimSpace(body.Profile)
|
||||
if profile == "" {
|
||||
profile = "standard"
|
||||
}
|
||||
name := taskDisplayName(target, "", "")
|
||||
if strings.TrimSpace(body.DisplayName) != "" {
|
||||
name = body.DisplayName
|
||||
}
|
||||
// Append profile tag.
|
||||
name = fmt.Sprintf("%s · %s", name, profile)
|
||||
|
||||
if target == "nvidia-bench-power" && parallelGPUs {
|
||||
writeError(w, http.StatusBadRequest, "power / thermal fit benchmark uses sequential or ramp-up modes only")
|
||||
return
|
||||
}
|
||||
resolved, err := expandSelectedGPUIndices(gpus, body.GPUIndices, body.ExcludeGPUIndices)
|
||||
if err != nil {
|
||||
writeError(w, http.StatusBadRequest, err.Error())
|
||||
return
|
||||
}
|
||||
if len(resolved) < 2 {
|
||||
// Fall through to normal single-task path.
|
||||
rampUp = false
|
||||
} else {
|
||||
now := time.Now()
|
||||
rampRunID := fmt.Sprintf("ramp-%s", now.UTC().Format("20060102-150405"))
|
||||
var allTasks []*Task
|
||||
for step := 1; step <= len(resolved); step++ {
|
||||
subset := resolved[:step]
|
||||
stepName := fmt.Sprintf("%s · ramp %d/%d · GPU %s", name, step, len(resolved), formatGPUIndexList(subset))
|
||||
t := &Task{
|
||||
ID: newJobID("benchmark-nvidia"),
|
||||
Name: stepName,
|
||||
Target: "nvidia-benchmark",
|
||||
Priority: defaultTaskPriority("nvidia-benchmark", taskParams{}),
|
||||
Status: TaskPending,
|
||||
CreatedAt: now,
|
||||
params: taskParams{
|
||||
GPUIndices: append([]int(nil), subset...),
|
||||
SizeMB: body.SizeMB,
|
||||
BenchmarkProfile: body.Profile,
|
||||
RunNCCL: runNCCL && step == len(resolved),
|
||||
ParallelGPUs: true,
|
||||
RampStep: step,
|
||||
RampTotal: len(resolved),
|
||||
RampRunID: rampRunID,
|
||||
DisplayName: stepName,
|
||||
},
|
||||
|
||||
if rampUp && len(body.GPUIndices) > 1 {
|
||||
// Ramp-up mode: resolve GPU list, then create one task per prefix
|
||||
// [gpu0], [gpu0,gpu1], ..., [gpu0,...,gpuN-1], each running in parallel.
|
||||
gpus, err := apiListNvidiaGPUs(h.opts.App)
|
||||
if err != nil {
|
||||
writeError(w, http.StatusBadRequest, err.Error())
|
||||
return
|
||||
}
|
||||
resolved, err := expandSelectedGPUIndices(gpus, body.GPUIndices, body.ExcludeGPUIndices)
|
||||
if err != nil {
|
||||
writeError(w, http.StatusBadRequest, err.Error())
|
||||
return
|
||||
}
|
||||
if len(resolved) < 2 {
|
||||
// Fall through to normal single-task path.
|
||||
rampUp = false
|
||||
} else {
|
||||
now := time.Now()
|
||||
rampRunID := fmt.Sprintf("ramp-%s", now.UTC().Format("20060102-150405"))
|
||||
var allTasks []*Task
|
||||
for step := 1; step <= len(resolved); step++ {
|
||||
subset := resolved[:step]
|
||||
stepName := fmt.Sprintf("%s · ramp %d/%d · GPU %s", name, step, len(resolved), formatGPUIndexList(subset))
|
||||
t := &Task{
|
||||
ID: newJobID("bee-bench-nvidia"),
|
||||
Name: stepName,
|
||||
Target: target,
|
||||
Priority: defaultTaskPriority(target, taskParams{}),
|
||||
Status: TaskPending,
|
||||
CreatedAt: now,
|
||||
params: taskParams{
|
||||
GPUIndices: append([]int(nil), subset...),
|
||||
SizeMB: body.SizeMB,
|
||||
BenchmarkProfile: body.Profile,
|
||||
RunNCCL: runNCCL && step == len(resolved),
|
||||
ParallelGPUs: true,
|
||||
RampStep: step,
|
||||
RampTotal: len(resolved),
|
||||
RampRunID: rampRunID,
|
||||
DisplayName: stepName,
|
||||
},
|
||||
}
|
||||
allTasks = append(allTasks, t)
|
||||
}
|
||||
allTasks = append(allTasks, t)
|
||||
for _, t := range allTasks {
|
||||
globalQueue.enqueue(t)
|
||||
}
|
||||
writeTaskRunResponse(w, allTasks)
|
||||
return
|
||||
}
|
||||
for _, t := range allTasks {
|
||||
globalQueue.enqueue(t)
|
||||
}
|
||||
writeTaskRunResponse(w, allTasks)
|
||||
}
|
||||
|
||||
// For non-ramp tasks append mode tag.
|
||||
if parallelGPUs {
|
||||
name = fmt.Sprintf("%s · parallel", name)
|
||||
} else {
|
||||
name = fmt.Sprintf("%s · sequential", name)
|
||||
}
|
||||
|
||||
params := taskParams{
|
||||
GPUIndices: body.GPUIndices,
|
||||
ExcludeGPUIndices: body.ExcludeGPUIndices,
|
||||
SizeMB: body.SizeMB,
|
||||
BenchmarkProfile: body.Profile,
|
||||
RunNCCL: runNCCL,
|
||||
ParallelGPUs: parallelGPUs,
|
||||
DisplayName: body.DisplayName,
|
||||
}
|
||||
tasks, err := buildNvidiaTaskSet(target, defaultTaskPriority(target, params), time.Now(), params, name, h.opts.App, "bee-bench-nvidia")
|
||||
if err != nil {
|
||||
writeError(w, http.StatusBadRequest, err.Error())
|
||||
return
|
||||
}
|
||||
for _, t := range tasks {
|
||||
globalQueue.enqueue(t)
|
||||
}
|
||||
writeTaskRunResponse(w, tasks)
|
||||
}
|
||||
}
|
||||
|
||||
// For non-ramp tasks append mode tag.
|
||||
if parallelGPUs {
|
||||
name = fmt.Sprintf("%s · parallel", name)
|
||||
} else {
|
||||
name = fmt.Sprintf("%s · sequential", name)
|
||||
}
|
||||
|
||||
params := taskParams{
|
||||
GPUIndices: body.GPUIndices,
|
||||
ExcludeGPUIndices: body.ExcludeGPUIndices,
|
||||
SizeMB: body.SizeMB,
|
||||
BenchmarkProfile: body.Profile,
|
||||
RunNCCL: runNCCL,
|
||||
ParallelGPUs: parallelGPUs,
|
||||
DisplayName: body.DisplayName,
|
||||
}
|
||||
tasks, err := buildNvidiaTaskSet("nvidia-benchmark", defaultTaskPriority("nvidia-benchmark", params), time.Now(), params, name, h.opts.App, "benchmark-nvidia")
|
||||
if err != nil {
|
||||
writeError(w, http.StatusBadRequest, err.Error())
|
||||
return
|
||||
}
|
||||
for _, t := range tasks {
|
||||
globalQueue.enqueue(t)
|
||||
}
|
||||
writeTaskRunResponse(w, tasks)
|
||||
func (h *handler) handleAPIBenchmarkNvidiaRun(w http.ResponseWriter, r *http.Request) {
|
||||
h.handleAPIBenchmarkNvidiaRunKind("nvidia-bench-perf").ServeHTTP(w, r)
|
||||
}
|
||||
|
||||
func (h *handler) handleAPISATStream(w http.ResponseWriter, r *http.Request) {
|
||||
|
||||
Reference in New Issue
Block a user