Performance regression on Windows due to removal of jemalloc

[iopq]

https://bitbucket.org/iopq/fizzbuzz-in-rust

Steps to reproduce:

1. Compile this with
   rustup install nightly-2016-01-20
cargo clean
rustup run nightly-2016-01-20 cargo bench

observe that you get nice performance:

running 2 tests
test bench_cowbuzz ... bench: 891 ns/iter (+/- 72)
test bench_fizzbuzz ... bench: 1,125 ns/iter (+/- 40)

2. Compile this with

rustup install nightly-2016-01-21
cargo clean
rustup run nightly-2016-01-21 cargo bench

observe that you get a performance regression:

test bench_cowbuzz ... bench: 1,071 ns/iter (+/- 42)
test bench_fizzbuzz ... bench: 1,542 ns/iter (+/- 101)

rustc -vV gives

commit-hash: 7dce32e
commit-date: 2016-01-20
host: i686-pc-windows-gnu
release: 1.7.0-nightly

3. The performance is worse in every version including the current nightly, 2016-06-12 gives

test bench_cowbuzz ... bench: 996 ns/iter (+/- 31)
test bench_fizzbuzz ... bench: 1,420 ns/iter (+/- 50)

[durka]

Seems to be within the error bars on OSX.

[iopq]

I found that it's my dependency that has the performance regression:

https://github.com/iopq/monoid

running the bench on 2016-01-20 gives

test bench_cowstring ... bench: 71 ns/iter (+/- 4)
test bench_cowstring_ms2ger ... bench: 66 ns/iter (+/- 5)

running the same thing on 2016-01-21 gives

test bench_cowstring ... bench: 130 ns/iter (+/- 5)
test bench_cowstring_ms2ger ... bench: 122 ns/iter (+/- 4)

the performance hasn't recovered since

all these tests are doing is string concatenation

#[bench]
fn bench_cowstring(b: &mut test::Bencher) {
    b.iter(|| {
        let fizz = Into::<Cow<_>>::into("Fizz").into_owned();
        let buzz = Into::<Cow<_>>::into("Buzz");
        Into::<Cow<_>>::into(fizz + &*buzz)
    });
}

#[bench]
fn bench_cowstring_ms2ger(b: &mut test::Bencher) {
    b.iter( || {  
            let mut fizz = Into::<Cow<_>>::into("Fizz").into_owned();
            let buzz = Into::<Cow<_>>::into("Buzz");
            fizz.push_str(&*buzz);
            Into::<Cow<_>>::into(fizz);
        }
    );
}

[TimNN]

The reduced test case does also not reproduce the problem on OS X.

Looking at the changes, this regression may have been caused by disabling jemalloc on windows-gnu in #30985 (edit: although it should have been re-enabled as far as I can tell, so if this was the cause, things should be fine on current nightlies).

[ollie27]

This does look like it's because jemalloc was disabled. On the 2016-09-08 nightly for x86_64-pc-windows-gnu I get:

test bench_cowstring        ... bench:         157 ns/iter (+/- 8)
test bench_cowstring_ms2ger ... bench:         155 ns/iter (+/- 11)

Adding #![feature(alloc_jemalloc)] and extern crate alloc_jemalloc; gives:

test bench_cowstring        ... bench:          85 ns/iter (+/- 3)
test bench_cowstring_ms2ger ... bench:          84 ns/iter (+/- 2)

Compared with the 2016-01-20 nightly:

test bench_cowstring        ... bench:          90 ns/iter (+/- 3)
test bench_cowstring_ms2ger ... bench:          91 ns/iter (+/- 6)

And the 2016-01-21 nightly:

test bench_cowstring        ... bench:         158 ns/iter (+/- 4)
test bench_cowstring_ms2ger ... bench:         156 ns/iter (+/- 2)

[brson]

Libs team needs to explain the path forward for re-gaining this perf loss. We're unlikely to add jemalloc back by default here.

[brson]

Right now the intended path forward to regain this perf is by stabilizing the global allocator API, and publishing the jemalloc crate to crates.io.