Perf tools updates and beyond
summary
Adds memory-oriented output and code annotations to data type profiling and improves instruction-pointer tracking for pointer arithmetic; current implementation maps samples to DWARF types on x86 and PowerPC but needs C++/Rust and broader arch support. Implements latency profiling that divides sample weight by measured parallelism using scheduler context-switch records and synthetic events (example: CC1 drops from 80% to 54% in latency view). Introduces deferred unwinding for user stacks, JSON event/metric descriptions, Python scripting, 10 ms lock-delay injection, and a guest-host ring-buffer proposal with correctness and PMU-access limitations.
AI-written summary · about
chapters (9)
- 0:00Introduction and overview
- 1:08Data type profiling updates
- 11:32Latency profiling explained
- 21:33Default unwinding support
- 33:18Event and metric JSON descriptions
- 36:34Lock contention profiling
- 40:31System core tracing
- 41:56Guest machine profiling
- 44:08Python support in Perf
related talks
37mAdvanced Performance Profiling Using Perf Tools
45mPerformance lab. The power of the perf tool!
16mBoost workload performance and get more done optimizing native code
36mDeveloping Continuous eBPF Profiler to look Beneath the Kernel to Beyond the Clouds
22mEnhancing RISC-V Performance Analysis Infrastructure
48mProfiling data structures
01h 02mObject Introspection: A C++ Memory Profiler
59mBeyond pprof, Hardware Performance Counters