主页 / 问题 / 598067
Accepted
Даниил Колесниченко
Asked:2020-12-01 16:38:31 +0000 UTC2020-12-01 16:38:31 +0000 UTC

为什么并行映射比顺序映射慢几个数量级？

  • 772

我想在 Rust 中试验并行性，特别是 - 比较map标准库中的串行和rayon中的并行map。这是基准代码：

#![feature(test)]

extern crate test;
extern crate rand;
extern crate rayon;

#[cfg(test)]
mod tests {
    use rayon::prelude::*;
    use test::Bencher;

    #[bench]
    fn iter_test(b: &mut Bencher) {
        let vec: Vec<i64> = (0..20000).collect();

        b.iter(|| -> i64 {
            vec.iter().map(|x| x * x).sum()
        });
    }

    #[bench]
    fn par_iter_test(b: &mut Bencher) {
        let vec: Vec<i64> = (0..20000).collect();

        b.iter(|| -> i64 {
            vec.par_iter().map(|x| x * x).sum()
        });
    }
}

结果：

running 2 tests
test tests::iter_test     ... bench:       8,061 ns/iter (+/- 2,606)
test tests::par_iter_test ... bench:   1,915,544 ns/iter (+/- 40,786,648)

为什么并行map这么慢？当然，我知道有些资源花在了并行化上，但没那么多。map编译器能以某种方式优化吗？那么如何编写一个基准，在这个基准上可以感受到并行的优势map？

PS 我试着传递一个map递归阶乘——结果差别不大，但顺序map的还是更快。

распараллеливание

1 个回答

  1. Best Answer

    因为在这种情况下，并行化的开销比有效负载高几个数量级。

    在我的机器上（Intel(R) Core(TM) i7-4770 CPU @ 3.40GHz），并行版本的性能开始超过 16,000,000 个值，超过 32,000,000 个值达到约 25%，并且没有进一步增长

    mkpankov@mkpankov-OptiPlex-9020 /tmp/rtest [master *]
± % ➜  cargo bench # 16000000
   Compiling rtest v0.1.0 (file:///tmp/rtest)
    Finished release [optimized] target(s) in 0.70 secs
     Running target/release/deps/rtest-14d3cc84caf097ae

running 2 tests
test tests::iter_test     ... bench:   8,749,482 ns/iter (+/- 465,264)
test tests::par_iter_test ... bench:   7,700,212 ns/iter (+/- 2,354,146)

test result: ok. 0 passed; 0 failed; 0 ignored; 2 measured


mkpankov@mkpankov-OptiPlex-9020 /tmp/rtest [master *]
± % ➜  cargo bench # 32000000
   Compiling rtest v0.1.0 (file:///tmp/rtest)
    Finished release [optimized] target(s) in 0.70 secs
     Running target/release/deps/rtest-14d3cc84caf097ae

running 2 tests
test tests::iter_test     ... bench:  17,497,810 ns/iter (+/- 765,301)
test tests::par_iter_test ... bench:  13,826,826 ns/iter (+/- 1,615,362)

test result: ok. 0 passed; 0 failed; 0 ignored; 2 measured


mkpankov@mkpankov-OptiPlex-9020 /tmp/rtest [master *]
± % ➜  cargo bench # 64000000
   Compiling rtest v0.1.0 (file:///tmp/rtest)
    Finished release [optimized] target(s) in 0.71 secs
     Running target/release/deps/rtest-14d3cc84caf097ae

running 2 tests
test tests::iter_test     ... bench:  34,855,214 ns/iter (+/- 1,047,213)
test tests::par_iter_test ... bench:  27,347,082 ns/iter (+/- 236,451)

test result: ok. 0 passed; 0 failed; 0 ignored; 2 measured
    • 2

相关问题