# Rustでパフォーマンス計測 RustのLT会 Shinjuku.rs #7 @FORCIA --- # About Me * @11Takanori * 仕事 * Kotlin, Rust, Python, Dart * Microservices --- # 話すこと - Rustで作られたツールで計測 - Rustで作られたアプリケーションを計測 --- # Rustで作られたツールで計測 --- # Welle * REST APIを計測するためのコマンドラインツール * cargo install welleでインストール --- ``` $ welle http://localhost:8080 --num-requests 10000 --concurrent-requests 10 Total Requests: 10000 Concurrency Count: 10 Total Completed Requests: 10000 Total Errored Requests: 0 Total 5XX Requests: 0 Total Time Taken: 960.743592ms Avg Time Taken: 96.074µs Total Time In Flight: 8.855463488s Avg Time In Flight: 885.546µs Percentage of the requests served within a certain time: 50%: 588.859µs 66%: 905.839µs 75%: 1.245908ms 80%: 1.554109ms 90%: 2.937985ms 95%: 5.523753ms 99%: 11.714656ms 100%: 13.083938ms ``` --- # Drill * REST APIを計測するためのコマンドラインツール * YAMLで記述 * cargo install drillでインストール --- ```yaml= threads: 4 base: "http://localhost:8080" iterations: 5 rampup: 2 plan: - name: index request: url: / ``` --- ``` $ drill --benchmark benchmark.yml --stats Threads 4 Iterations 5 Rampup 2 Base URL http://localhost:8080 index http://localhost:8080/ 200 OK 1ms index http://localhost:8080/ 200 OK 0ms index http://localhost:8080/ 200 OK 0ms index http://localhost:8080/ 200 OK 3ms index http://localhost:8080/ 200 OK 0ms index http://localhost:8080/ 200 OK 1ms ... index Total requests 20 index Successful requests 20 index Failed requests 0 index Median time per request 1ms index Average time per request 1ms index Sample standard deviation 2ms Concurrency Level 4 Time taken for tests 0.1 seconds Total requests 20 Successful requests 20 Failed requests 0 Requests per second 157.01 [#/sec] Median time per request 1ms Average time per request 1ms Sample standard deviation 2ms ``` --- # Rustで作られたアプリケーションを計測 --- # cargo bench * cargo benchで計測 * nightlyのみ対応 * stableとbetaの場合、[bencher](https://crates.io/crates/bencher)というcrateで似たようなことができる --- 計測対象のコード ```rust= fn fib(n: u32) -> u32 { if n <= 1 { n } else { fib(n - 2) + fib(n - 1) } } ``` --- ```rust= #[cfg(test)] mod tests { use super::*; use test::Bencher; #[bench] fn bench_fib(b: &mut Bencher) { b.iter(|| { let n = test::black_box(10); fib(n) // 10をfibに直接渡した場合、計算されない }); } } ``` --- ``` $ cargo bench running 1 test test tests::bench_fib ... bench: 287 ns/iter (+/- 25) test result: ok. 0 passed; 0 failed; 0 ignored; 1 measured; 0 filtered out ``` --- # flame * Cargo.tomlのdependenciesにflameを追加 * 計測したい箇所のコードを修正する必要がある --- ```ruby= fn main() { flame::start("read file"); let _x = read_a_file(); // 処理に100msかかる flame::end("read file"); let _y = flame::span_of("database query", || query_database()); // 処理に200msかかる let _z = { let _guard = flame::start_guard("cpu-heavy calculation"); cpu_heavy_operations_1(); // 処理に300msかかる flame::note("something interesting happened", None); cpu_heavy_operations_2(); // 処理に100msかかる }; flame::dump_json(&mut File::create("flame.json").unwrap()).unwrap(); } ``` --- ```json= $ jq . flame.json [ { "id": 139687886300416, "name": "main", "spans": [ { "name": "read file", "start_ns": 6641, "end_ns": 100083792, "delta": 100077151, "depth": 0, "children": [], "notes": [] }, { "name": "database query", "start_ns": 100087006, "end_ns": 300203617, "delta": 200116611, "depth": 0, "children": [], "notes": [] }, { "name": "cpu-heavy calculation", "start_ns": 300211351, "end_ns": 700414510, "delta": 400203159, "depth": 0, "children": [], "notes": [ { "name": "something interesting happened", "description": null, "instant": 600323628 } ] } ] } ] ``` --- # まとめ * Rustで作られた計測ツールがいくつかある * コードを修正すれば、内部の処理の計測も簡単 --- # おまけ --- # flamegraph * cargo install flamegraph * perfを外部プロセスとして実行するのでperfのインストールが必要 * Ubuntu18.04なら `sudo apt install linux-tools-4.15.0-66-generic` * CPUの処理時間を計測 * コードの修正不要 --- 下の内容をCargo.tomlに書いてcargo flamegraphを実行すれば、flamegraph.svgが生成される ```toml [profile.release] debug = true ``` ---