Benchmarking: assigning/affinity of CPU caches

[peter-lyons-kehl]

Any tips/starters on benchmarking/assigning/limiting CPU cache usage (on x64 Linux with systemd, initially with XOrg and on Manjaro/Arch)? But, while still using GNOME/similar + Firefox/VS Code.

(Without using cgroups directly, if possible.) Any GitHub/other CI tips, and local Mac OS tips, are welcome (but initially focused on local Linux with XOrg).

To partially isolate L1+L2 caches, we could use taskset or systemd-run to attach Firefox/VS Code to one hardware core, and the benchmarked app (single process, single threaded) to another hardware core. But, that doesn't handle L3.

And/or, any easy ways to restrict cores & CPU caches of GNOME/window manager itself:

• not starting GNOME login automatically at boot, or
• applying systemd and/or cgroups to existing GNOME/window manager processes,
• any other window managers,
• Wayland vs. XOrg
• other ways?

(OK to link/include C/x64 assembly, as far as it's friendly to LLVM. OK to limit it to pre-NUMA Intel generations only, at least initially).
(OK with unsafe and/or Rust's own nightly-only benchmarking framework (instead of Criterion). As of May 2024, benches/shared/lib_benches.rs uses Criterion, but we can move away from it easily.)