Profiling

Most standard profilers do not work with Motor OS yet, but a Motor OS program can be instrumented with the tracing crate and its trace turned into a flame graph on the host, which is enough to find, for example, the causes of high CPU contention.

Instrument the code

• Add tracing to the crates you want to instrument: tracing = { version = "0.1", default-features = false, features = ["attributes"] }.
• Add use tracing::instrument; to the modules and put #[instrument(skip_all)] on the functions you want to see (docs).
• Wrap code blocks inside functions with tracing::trace_span!("name").in_scope(|| { /* ... */ }); (docs).
• Add to the binary's Cargo.toml:

  tracing = "0.1"
  tracing-subscriber = "0.3"
  tracing-flame = "0.2"

• In the binary, wrap the region you want to profile. Write the trace somewhere an interactive program may create files, such as /user/tmp:

  use tracing_flame::FlameLayer;
  use tracing_subscriber::prelude::*;
  
  const FNAME: &str = "/user/tmp/profile.folded";
  let (flame_layer, guard) = FlameLayer::with_file(FNAME).unwrap();
  tracing_subscriber::registry().with(flame_layer).init();
  
  tracing::info_span!("profiling_root").in_scope(|| {
      /* the code you want to profile */
  });
  core::mem::drop(guard); // writes the trace to the file

Collect and visualize

• Run the binary in Motor OS.
• Copy the profile to the Linux host: scp -P 2222 -o IdentitiesOnly=yes -i test.key [email protected]:/user/tmp/profile.folded .
• cargo install inferno
• cat profile.folded | inferno-flamegraph > profile.svg

profile.svg is the CPU profile flame graph. profile.folded is a text file; you may need to edit it by hand or adjust inferno's options to make the graph readable.