perf_event metric source

Performance counter source for joule-profiler. using the Linux perf_event subsystem. This crate implements MetricSource from joule-profiler-core and collects hardware and software performance counters (CPU cycles, instructions, cache misses, branch mispredictions…) via the perf_event_open(2) syscall, per phase.

Requirements

Adjusting perf_event_paranoid

# Check current value
cat /proc/sys/kernel/perf_event_paranoid
 
# Allow per-process counters for unprivileged users (temporary)
sudo sysctl -w kernel.perf_event_paranoid=1
 
# Persistent
echo 'kernel.perf_event_paranoid=1' | sudo tee /etc/sysctl.d/99-perf.conf
sudo sysctl --system

Scope

By default, Joule Profiler attaches perf_event counters to the monitored process only (per-process mode, via the pid), following it and any children it spawns across every CPU it runs on.

Alternatively, counters can be scoped to a cgroup v2 instead of the process's pid, by setting cgroup_name in the source configuration:

[sources.perf]
cgroup_name = "my-cgroup" # use "parent/child" for a nested cgroup
#cgroup_root = "/my/cgroup/hierarchy" # default is /sys/fs/cgroup

This tracks every process inside the cgroup rather than a single pid and its children, it can be useful for containers profiling, and lets counters be opened before the profiled process is spawned (the cgroup must already exist).