Processor tracing in perf
summary
Intel Processor Trace (Intel PT) records every branch in a highly compressed packet stream (TNT bits, PSB sync, timestamps), decodes with perf userspace via the --itrace option, and lets perf synthesize instructions or call traces; BTS stores branches as BTS 24+8 bytes records and LBR keeps a small history (e.g. LBR depth 32), LBR incurs ~4x slowdown in one benchmark (baseline 58s), PT has much lower overhead and supports snapshot mode to dump traces on signal.
AI-written summary · about
chapters (9)
- 0:00Start
- 0:22Processor Tracing
- 1:46Branch Instructions
- 2:29Conditional Jumps
- 4:50Translate the Hex Dump
- 9:23Perf Report
- 24:40Call Graph
- 24:52Trace the Kernel
- 28:17Snapshot Mode
related talks
59mThe Magic of a Trace
46mbpftrace internals
16mProcessor Trace in a Holistic World
37mKernel and userspace tracing with LTTng and friends
55mRecent Advances in Linux Tracing
34mTrace-Profiling: A New Way About How to Track Application Behavior
30mPerformance insights into eBPF step by step
52mTracking Down What Happened: A Comparison of Tracing Solutions