Beyond pprof, Hardware Performance Counters
summary
perfgo collects PMU data via Linux perf and converts it into pprof so you can map perfgo profiles to Go source; dTLB instrumentation shows Grafana Labs cut TLB misses by moving to huge pages but it costs ~7% more RSS and yields no measurable runtime gain for that workload using dTLB cache miss counters; focusing on branches, adding three lexer fast paths drops a hot function from ~17.5% to ~9% CPU and shrinks branch misses from the hot path (observed 14% branch-miss rate on random input); note many cloud VMs hide PMUs and transparent huge pages have allocation and latency tradeoffs.
AI-written summary · about
chapters (13)
- 0:00Hardware performance counters
- 1:55CPU architecture evolution
- 8:17Cache locality and prefetching
- 10:26Hardware counters and PMU
- 11:31Prometheus workload analysis
- 16:30TLB optimization with huge pages
- 23:14Branch prediction misses
- 33:20Benchmarking methodology
- 36:53Perf Go profiling tool
- 39:00Branch optimization demo
- 44:44Kubernetes profiling with Perf Go
- 46:43Summary
- 48:28Q&A
