#low-overhead #multi-threading #applications #events #frequency #high

micrometer

Profiling for fast, high frequency events in multithreaded applications with low overhead

9 releases

0.2.7 Oct 24, 2023
0.2.6 Jul 13, 2023
0.2.3 Apr 20, 2023
0.2.0 Mar 20, 2023
0.1.0 Feb 27, 2023

#66 in Profiling


Used in 2 crates

Apache-2.0

36KB
754 lines

Micrometer

Crates.io Documentation

Profiling for fast, high frequency events in multithreaded applications with low overhead

Important: enabling data collection

By default every measure is a no-op, to measure and consume measures, enable the enable feature. This is done to allow libs to instrument their code without runitme costs if no one is using the measures.

Definitions

  • Measurement: a measurement is the timing data gathered for a single event.
  • Record: a record is the retrieval format for measurements. Records may be 1-to-1 with respects to measurements, however, if perforation is enabled (default), as measurements increase in numbers, a single record may be the aggregate of multiple measurements.
  • Location: a location is a region of interest of the program. All measurements start from a location. Each measure is associated with the name of the location it originated from.
  • Thread location: as micrometer is intended for multithreaded environments, each location may be relevant for multiple threads, the location in the context of a specific thread is referred to as thread location
  • Span: a span is a guard object that measures the time between when it was created and when it was dropped (unless it was disarmed). Each span will produce one measurement
  • Track: a track is the struct responsible of handling all measurements originating from a thread location. It is used to create spans or manually record the duration of an event.
  • TrackPoint: while a track maps to the concept of thread location (as a single track can only be owned by one thread), a track point maps to the concept of location. The track point struct is a lazily allocated, thread local handle to multiple tracks that are associated with the same name. Usually, track points are used as static variables, using the get_or_init method to get the track for the active thread.

Examples

Measuring the duration of a loop

for _ in 0..100 {
    // Define a `TrackPoint` named "loop_duration", get (or init) the `Track` for the
    // current thread, then create a span and assign it to a new local variable called
    // `loop_duration`
    micrometer::span!(loop_duration);

    std::hint::black_box("do something");
    // `loop_duration` is automatically dropped recording the measure
}
// Print a summary of the measurements
micrometer::summary();

Measuring the duration of a loop, threaded

std::thread::scope(|s| {
    for t in 1..=4 {
        s.spawn(move || {
            for _ in 0..(10 * t) {
                micrometer::span!(); // Name automatically assigned to source file and line
                std::hint::black_box("do something");
            }
        });
    }
});
// Print a summary of the measurements
micrometer::summary();
// Print a summary of the measurements, aggregating measures for the same location
micrometer::summary_grouped();

Measuring the duration of an expression

// This works like the `dbg!` macro, allowing you to transparently wrap an expression:
// a span is created, the expression is executed, then the span is closed and the result
// of the expression is passed along
let a = micrometer::span!(5 * 5, "fives_sq");
let b = micrometer::span!(a * a); // Name automatically assigned to source file and line
assert_eq!(a, 25);
assert_eq!(b, 25 * 25);

Measuring a code segment

let a = 5;
micrometer::span!(guard, "a_sq");
let b = a * a;
drop(guard); // Measurement stops here
let c = b * a;

Dependencies

~0.4–5.5MB
~12K SLoC