Linux perf tool metrics
summary
Linux perf now ships hundreds of metrics (Valkyrie contributions) encoded as JSON and compiled into perf via create_perf_json.py, letting metric expressions combine raw PMU events into readable values. Top‑down microarchitecture (TMA) metrics enforce parent/child sums (level‑1 totals 100%), but thresholds can add events and force multiplexing; EBS (event‑based sampling) is buggy so TMA defaults are conservative. A new grouping algorithm uses per‑PMU counter limits to reduce multiplexing (RFC posted), and timed PS retire‑latency sampling is being integrated for more accurate metrics.
AI-written summary · about
chapters (9)
- 0:00Introduction to perf metrics
- 6:51Who defines metric formulas
- 8:58Top-down microarchitecture analysis
- 16:31EBS mode controversy
- 18:57Hybrid processor metrics
- 20:22Metric accuracy validation
- 21:49Event grouping challenges
- 33:56Grouping method Q&A
- 39:34Timed PS feature overview
related talks
59mBeyond pprof, Hardware Performance Counters
45mPerformance lab. The power of the perf tool!
29mPerf Support in Xen
45mPerf tools updates and beyond
35mIntegrated PCIe monitoring and tracing facilities
20mTintin: A Unified Hardware Performance Profiling Infrastructure to Uncover and Manage...
51mLinux Systems Performance - Architecture
24mlibperf: library for perf events monitoring