the #perf #documentation #tutorial, which explains -e cycles:u and the like.
on 02023-07-03with #perf you are measuring the kernel too by default; perf stat -e instructions:u ./cmd to only see user-space instructions gives more consistent results (varying by 0.01% instead of 10%, and within 0.1% of cachegrind) but still always different on every run.
how to use #perf and "tiptop" to measure #performance in terms of average instructions per clock with PMCs (performance monitoring counters).
on 02023-07-03Mozilla #rr supports record-replay backward stepping on fewer platforms than the #GDB #debugger (only amd64 and then only sometimes). But it supports recording the traces to disk. Also, apparently, it requires /proc/sys/kernel/perf_event_paranoid to be no more than 1 for its fast (non -n) mode (like #perf), and doesn’t work inside Docker on the Kudu server, but does work inside KVM. #reverse-debugging
notes from 02009 on #tracing with #Linux tracepoints, including the use of #perf. On Ubuntu you need to install, for example, linux-tools-5.8.0-40-generic if you’re running linux-image-5.8.0-40-generic.
on 02021-02-03how to use #Linux kernel #tracing tracepoints effectively and analyze #performance with #perf
on 02021-02-03Examples of #profiling #performance with "perf_events", sometimes called "perf".
on 02016-07-24