# Benchmarking `factor` The benchmarks for `factor` are located under `tests/benches/factor` and can be invoked with `cargo bench` in that directory. They are located outside the `uu_factor` crate, as they do not comply with the project's minimum supported Rust version, *i.e.* may require a newer version of `rustc`. ## Microbenchmarking deterministic functions We currently use [`criterion`] to benchmark deterministic functions, such as `gcd` and `table::factor`. However, microbenchmarks are by nature unstable: not only are they specific to the hardware, operating system version, etc., but they are noisy and affected by other tasks on the system (browser, compile jobs, etc.), which can cause `criterion` to report spurious performance improvements and regressions. This can be mitigated by getting as close to [idealized conditions][lemire] as possible: - minimize the amount of computation and I/O running concurrently to the benchmark, *i.e.* close your browser and IM clients, don't compile at the same time, etc. ; - ensure the CPU's [frequency stays constant] during the benchmark ; - [isolate a **physical** core], set it to `nohz_full`, and pin the benchmark to it, so it won't be preempted in the middle of a measurement ; - disable ASLR by running `setarch -R cargo bench`, so we can compare results across multiple executions. [`criterion`]: https://bheisler.github.io/criterion.rs/book/index.html [lemire]: https://lemire.me/blog/2018/01/16/microbenchmarking-calls-for-idealized-conditions/ [isolate a **physical** core]: https://pyperf.readthedocs.io/en/latest/system.html#isolate-cpus-on-linux [frequency stays constant]: ... ### Guidance for designing microbenchmarks *Note:* this guidance is specific to `factor` and takes its application domain into account; do not expect it to generalize to other projects. It is based on Daniel Lemire's [*Microbenchmarking calls for idealized conditions*][lemire], which I recommend reading if you want to add benchmarks to `factor`. 1. Select a small, self-contained, deterministic component (`gcd` and `table::factor` are good examples): - no I/O or access to external data structures ; - no call into other components ; - behavior is deterministic: no RNG, no concurrency, ... ; - the test's body is *fast* (~100ns for `gcd`, ~10µs for `factor::table`), so each sample takes a very short time, minimizing variability and maximizing the numbers of samples we can take in a given time. 1. Benchmarks are immutable (once merged in `uutils`) Modifying a benchmark means previously-collected values cannot meaningfully be compared, silently giving nonsensical results. If you must modify an existing benchmark, rename it. 1. Test common cases We are interested in overall performance, rather than specific edge-cases; use **reproducibly-randomized inputs**, sampling from either all possible input values or some subset of interest. 1. Use [`criterion`], `criterion::black_box`, ... `criterion` isn't perfect, but it is also much better than ad-hoc solutions in each benchmark. ## Wishlist ### Configurable statistical estimators `criterion` always uses the arithmetic average as estimator; in microbenchmarks, where the code under test is fully deterministic and the measurements are subject to additive, positive noise, [the minimum is more appropriate][lemire]. ### CI & reproducible performance testing Measuring performance on real hardware is important, as it relates directly to what users of `factor` experience; however, such measurements are subject to the constraints of the real-world, and aren't perfectly reproducible. Moreover, the mitigation for it (described above) isn't achievable in virtualized, multi-tenant environments such as CI. Instead, we could run the microbenchmarks in a simulated CPU with [`cachegrind`], measure execution “time” in that model (in CI), and use it to detect and report performance improvements and regressions. [`iai`] is an implementation of this idea for Rust. [`cachegrind`]: https://www.valgrind.org/docs/manual/cg-manual.html [`iai`]: https://bheisler.github.io/criterion.rs/book/iai/iai.html ### Comparing randomized implementations across multiple inputs `factor` is a challenging target for system benchmarks as it combines two characteristics: 1. integer factoring algorithms are randomized, with large variance in execution time ; 1. various inputs also have large differences in factoring time, that corresponds to no natural, linear ordering of the inputs. If (1) was untrue (i.e. if execution time wasn't random), we could faithfully compare 2 implementations (2 successive versions, or `uutils` and GNU) using a scatter plot, where each axis corresponds to the perf. of one implementation. Similarly, without (2) we could plot numbers on the X axis and their factoring time on the Y axis, using multiple lines for various quantiles. The large differences in factoring times for successive numbers, mean that such a plot would be unreadable.