rayon_1_0_0::sort perf regression

[meven]

According to the work done on lolbench there was a perf regression on
rayon_1_0_0::sort::demo_merge_sort_descending and rayon_1_0_0::sort::par_sort_descending between nightly 2018-11-03 and 2018-11-04.

The benchmarks grew 100%, and close to 200 % in branch_instructions / iteration.

cc @anp

https://lolbench.rs/benchmarks/rayon-1-0-0-sort-demo-merge-sort-descending.html
https://lolbench.rs/benchmarks/rayon-1-0-0-sort-par-sort-descending.html

[pnkfelix]

tagging as P-high for the initial investigation of the scope of the performance regression.

[nikic]

Commit range is 8b09631...04fdb44, which notably includes the jemalloc removal. Most likely this is just an allocator regression and can be recovered by explicitly enabling jemalloc.

[pnkfelix]

I can observe the ~2.0x slowdown regression on my Linux desktop.

My mac does not exhibit it nearly as cleanly.

I'm going to see if switching on jemallocator manually (by adding it as a #[global_allocator] to rayon-demo's main.rs) eliminates the regression.

• Update: yes, turning on #[global_allocator] static ALLOC: jemallocator::Jemalloc = jemallocator::Jemalloc; eliminates the regression.

[pnkfelix]

Okay so I think that confirms the hypothesis from @nikic

You can see from the description on PR #55238 that switching from jemalloc to glibc alloc was expected to cause regressions. That included the note that you would get that performance back (and then some) by switching to jemalloc-sys from crates.io. The Rust stdlib don't do it by default, but as far as I can tell it is what we are recommending as the path for clients who want to maximize performance on this axis?

[pnkfelix]

In any case my personal recommendation is that Rayon should consider turning on jemalloc for its demo code.

Here is the necessary diff:

diff --git a/rayon-demo/Cargo.toml b/rayon-demo/Cargo.toml
index 897b5c8..e1aba34 100644
--- a/rayon-demo/Cargo.toml
+++ b/rayon-demo/Cargo.toml
@@ -5,6 +5,7 @@ authors = ["Niko Matsakis <niko@alum.mit.edu>"]
 publish = false

 [dependencies]
+jemallocator = "0.1.8"
 rayon = { path = "../" }
 cgmath = "0.16"
 docopt = "1"
diff --git a/rayon-demo/src/main.rs b/rayon-demo/src/main.rs
index 84619c0..e21d7b0 100644
--- a/rayon-demo/src/main.rs
+++ b/rayon-demo/src/main.rs
@@ -1,5 +1,9 @@
 #![cfg_attr(test, feature(test))]

+extern crate jemallocator;
+#[global_allocator]
+static ALLOC: jemallocator::Jemalloc = jemallocator::Jemalloc;
+
 use std::env;
 use std::io;
 use std::io::prelude::*;

[pnkfelix]

(I think there isn't much more to investigate here with respect to the scope of the regression nor its source. Downgrading from P-high to P-medium, and nominating for discussion at next compiler meeting with inclination to close as "not-a-bug"/"wont-fix")

[anp]

FWIW I hope to support measuring against multiple different allocators in lolbench in the future, just not there yet.