Add real-data duration estimates to benchmark and burn pages

- Add BenchmarkEstimated* constants to benchmark_types.go from _v8 logs
  (Standard Perf ~16 min, Standard Power Fit ~43 min, Stability Perf ~92 min)
- Update benchmark profile dropdown to show Perf / Power Fit timing per profile
- Add timing columns to Method Split table (Standard vs Stability per run type)
- Update burn preset labels to show "N min/GPU (sequential) or N min (parallel)"
- Clarify burn "one by one" description with sequential vs parallel scaling

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
2026-04-18 10:54:50 +03:00
parent bac89bb6e5
commit 51b721aeb3
2 changed files with 36 additions and 11 deletions

View File

@@ -43,6 +43,31 @@ const (
NvidiaBenchmarkProfileOvernight = "overnight"
)
// Estimated wall-clock durations for benchmark runs, derived from real _v8 logs.
// Rule: when changing profile phase durations in resolveBenchmarkProfile(),
// re-measure from actual task logs and update the constants here.
//
// Sources:
// - BenchmarkEstimatedPerfStandardSec: MLT v8.22 ramp 1-4: 927 s; xFusion v8.22 parallel 8GPU: 1080 s
// - BenchmarkEstimatedPerfStabilitySec: xFusion v8.22 ramp 1-8: 5532 s
// - BenchmarkEstimatedPerfOvernightSec: derived from profile phases (SteadySec=27000)
// - BenchmarkEstimatedPowerStandardSec: MLT v8.22 ramp 1-4: 2663 s; MSI v8.22 ramp 1-8: 2375 s
// - BenchmarkEstimatedPowerStabilitySec: xFusion v8.17/v8.22 ramp 1-8: 1977-2002 s
const (
// Performance Benchmark (bee-gpu-burn).
// Duration is per full ramp-up run (ramp 1→N) or per single parallel run.
// Sequential per-GPU mode scales approximately linearly.
BenchmarkEstimatedPerfStandardSec = 960 // ~16 min; ramp-up 1-4: 927 s, parallel 8GPU: 1080 s
BenchmarkEstimatedPerfStabilitySec = 5532 // ~92 min; ramp-up 1-8 measured
BenchmarkEstimatedPerfOvernightSec = 8 * 3600
// Power / Thermal Fit (dcgmi targeted_power binary-search calibration).
// Duration is for the full ramp-up run; individual steps vary with convergence speed.
BenchmarkEstimatedPowerStandardSec = 2600 // ~43 min; ramp 1-4: 2663 s, ramp 1-8: 2375 s
BenchmarkEstimatedPowerStabilitySec = 2000 // ~33 min; stability profile converges faster (longer steady → faster convergence)
BenchmarkEstimatedPowerOvernightSec = 3 * 3600
)
type NvidiaBenchmarkOptions struct {
Profile string
SizeMB int