bench

Overview ¶

Command bench is a thesis-grade measurement harness for bubblepprof's target-side profiling paths. It spawns a parameterized synthetic workload, drives N iterations of either memusage.Compute (in-process analysis) or bundle.CaptureSelf (external analyser target-side capture only) against the running process, and emits a JSON report with per-iteration measurements and aggregate statistics.

Captured per iteration:

• wall-clock latency of the selected operation
• max user-visible scheduling pause during the operation (heartbeat goroutine)
• Go heap allocated by the operation (HeapAlloc delta, TotalAlloc delta)
• process RSS before/after and per-call peak RSS delta on Linux (/proc/self/status VmRSS/VmHWM, optionally reset via clear_refs)
• matched_goroutines, reachable_bytes from the response in compute mode
• bundle_bytes emitted in bundle mode

Aggregates: mean, stddev, min, max, p50, p95, p99 per metric.

One iteration optionally runs under runtime/trace (-trace=path.trace) so the per-stage spans recorded by internal/memusage's trace.StartRegion calls can be inspected with `go tool trace` and cross-checked against the heartbeat-based STW estimate.