Replace jemalloc with mimalloc #92249

fee1-dead · 2021-12-24T10:48:41Z

supersedes #81782

rust-highfive · 2021-12-24T10:48:44Z

r? @matthewjasper

(rust-highfive has picked a reviewer for you, use r? to override)

dtolnay · 2021-12-24T12:13:00Z

@bors try

bors · 2021-12-24T12:13:08Z

⌛ Trying commit b90cfc8 with merge e7033aafc2d47cc9d3ebe315ec20e2b5a9e586da...

takkuumi · 2021-12-24T12:22:01Z

Why replace jemalloc?

AngelicosPhosphoros · 2021-12-24T12:23:21Z

config.toml.example

@@ -582,9 +582,9 @@ changelog-seen = 2
 # Map debuginfo paths to `/rust/$sha/...`, generally only set for releases
 #remap-debuginfo = false

-# Link the compiler against `jemalloc`, where on Linux and OSX it should
+# Link the compiler against `mimalloc`, where on Linux and OSX it should


Could we use mimalloc on Windows too? It is created by MS so it should be supported, probably.

https://github.com/microsoft/mimalloc#override-on-windows

AngelicosPhosphoros · 2021-12-24T12:28:27Z

@NateLing It can be faster, probably.

However, PR really lacks description yet.

takkuumi · 2021-12-24T13:17:51Z

@NateLing It can be faster, probably.

However, PR really lacks description yet.

In my program (a web server based on hyper and tower), wrk test result could be 270000 rps +-, but mimalloc rps is 260000+-

fee1-dead · 2021-12-24T13:50:40Z

In my program (a web server based on hyper and tower), wrk test result could be 270000 rps +-, but mimalloc rps is 260000+-

In the last PR the results were all green on perf.rlo. In general there is no best allocator and it depends on the type of the program,

AngelicosPhosphoros · 2021-12-24T13:52:31Z

In my program (a web server based on hyper and tower), wrk test result could be 270000 rps +-, but mimalloc rps is 260000+-

Did you disabled MiMallocs secure mode? It comes with cost and it is not necessary for correct Rust program.

Also, MiMalloc works on Windows too so we would be able to get +1 platform with better allocator in compiler.

bors · 2021-12-24T14:38:05Z

☀️ Try build successful - checks-actions
Build commit: e7033aafc2d47cc9d3ebe315ec20e2b5a9e586da (e7033aafc2d47cc9d3ebe315ec20e2b5a9e586da)

takkuumi · 2021-12-24T14:49:31Z

@AngelicosPhosphoros yep, i disabled MiMallocs secure mode.
@fee1-dead i agree you. i dont care about 10000 rps. i think there could be a guide doc, so user can choose the right malloc in rust program.

fee1-dead · 2021-12-26T06:54:12Z

Assuming @dtolnay wanted a perf run:

@rust-timer build e7033aafc2d47cc9d3ebe315ec20e2b5a9e586da

Regarding windows support: I will investigate how to do this in the next few days.

rust-timer · 2021-12-26T06:54:14Z

Queued e7033aafc2d47cc9d3ebe315ec20e2b5a9e586da with parent 59337cd, future comparison URL.

rust-timer · 2021-12-26T08:21:49Z

Finished benchmarking commit (e7033aafc2d47cc9d3ebe315ec20e2b5a9e586da): comparison url.

Summary: This change led to very large relevant improvements 🎉 in compiler performance.

Very large improvement in instruction counts (up to -10.3% on incr-unchanged builds of deep-vector)

If you disagree with this performance assessment, please file an issue in rust-lang/rustc-perf.

Benchmarking this pull request likely means that it is perf-sensitive, so we're automatically marking it as not fit for rolling up. While you can manually mark this PR as fit for rollup, we strongly recommend not doing so since this PR led to changes in compiler perf.

@bors rollup=never
@rustbot label: +S-waiting-on-review -S-waiting-on-perf -perf-regression

Mark-Simulacrum · 2021-12-26T12:56:58Z

FWIW it looks like the crates.io libmimalloc-sys is pretty out of date (2 years old), so while the perf results look good I'd suggest experimenting with a fork that brings mimalloc to a more recent release.

In general though results look pretty promising!

fee1-dead · 2021-12-26T13:12:23Z

it looks like the crates.io libmimalloc-sys is pretty out of date (2 years old)

You should be looking at https://crates.io/crates/libmimalloc-sys which was updated just a month ago which I believe is the latest release.

nikic · 2021-12-26T14:48:33Z

The perf results still show a 35% regression in memory usage.

fee1-dead · 2021-12-26T15:53:48Z

The perf results still show a 35% regression in memory usage.

Seems to be microsoft/mimalloc#393. They have suggested using v2.x but those versions are currently in beta and AFAIK there is no crate that currently package that major version. I'm going to make a fork and test it out.

jq-rs · 2022-01-05T17:50:22Z

compiler/rustc/src/main.rs

        // registers itself with the allocator's zone APIs in a ctor. However,
        // the linker doesn't seem to consider ctors as "used" when statically
        // linking, so we need to explicitly depend on the function.
        #[cfg(target_os = "macos")]
        {
            extern "C" {
-                fn _rjem_je_zone_register();
+                fn _mi_macos_override_malloc();


Thanks for the superseded PR, great to see this is progressing! Out of curiosity, I tried enabling debugs for libmimalloc-sys on macOS BigSur and I could not get the debug statistics increasing with env MIMALLOC_VERBOSE=1 MIMALLOC_SHOW_STATS=1 ./x.py build with this commit:

... heap stats: peak total freed current unit count normal: 0 0 0 0 ok huge: 0 0 0 0 ok giant: 0 0 0 0 ok total: 0 0 0 0 ok malloc req: 0 0 0 0 ok reserved: 0 0 0 0 ok committed: 0 0 0 0 ok reset: 0 0 0 0 ok touched: 0 0 0 0 ok segments: 0 0 0 0 ok -abandoned: 0 0 0 0 ok -cached: 0 0 0 0 ok pages: 0 0 0 0 ok -abandoned: 0 0 0 0 ok -extended: 0 -noretire: 0 mmaps: 0 commits: 0 threads: 0 0 0 0 ok searches: 0.0 avg mimalloc: using 1 numa regions numa nodes: 1 elapsed: 11.095 s process: user: 11.671 s, system: 0.478 s, faults: 100, rss: 223.9 MiB mimalloc: process done: 0x10df93e00 ...

Maybe something more is missing? This is where I got stuck earlier. I can try again if someone has these working, maybe something related to my debug-enable? On Linux however, these do show values when override works.

Hmm, I did not test this in macOS because I don't have a machine accessible. Windows is not supported because static linking cannot be used with override. Not sure what the next steps are...

The results look really good to me, so I think we should try to find a way to fix this, if possible. I am happy to test this, but I am not proficient with macOS either. Maybe @sfackler or @ehuss have ideas what to try next (as you fixed jemalloc for macOS earlier)?

Also, @thomcc may be interested to give new ideas.

You have to enable stats when building or they aren't tracked, (see: https://github.com/microsoft/mimalloc/blob/master/include/mimalloc-types.h#L32-L33). I don't remember if theres support in the mimalloc (or -sys) crates directly, but setting env CFLAGS="-DMI_STAT=1" or env CFLAGS="-DMI_STAT=2" when building should make cc do it, hopefully.

@thomcc Thanks, I checked this in detail. In my debug build I enable MI_DEBUG_FULL in CMakeLists.txt and this will enable MI_DEBUG=3definition. This then again will set MI_STAT=2 https://github.com/microsoft/mimalloc/blob/master/include/mimalloc-types.h#L421-L427. I cannot easily see what could go wrong, but to be sure, I'll try with a separate forced setting to MI_STAT.

If the statistics are valid, then mi_macos_override_malloc seems not to be called, perhaps due to similar reasons jemalloc originally did not work (#82423), where the linker is optimizing away some needed call paths. We would need to figure out what is still missing for mimalloc. My skills need sharpening here, so any hints on how to do it best are welcome :)

Do you have your integration code published anywhere? If not I can take a stab, but would like to avoid redundant work.

I think the mi_macos_override_malloc stuff is requires a little bit of extra work. most of the documentatoin mimalloc has about what's required to override on macos is a bit dated (or wrong) too (interposing happens by default, flat namespace should not be required).

Also, the zone override is not intended to be used on its own, but that's fixable.

@thomcc Thank you for taking a look! My current changes are just testing and stats debug related, which we discussed above. I do not have anything else outside this PR.

Taking more of a look, it's very unsurprising that you can't call _mi_macos_override_malloc directly, it's a static private function. It's also only included when mimalloc is built as a dynamic library (there are various reasons for this, but I think the main one is that mimalloc attempts interposing, and prior to macOS 10.12, this could only be done as a dynamic lib with some DYLD_ environment vars defined).

There are at least a few issues with the way rustc will need to use it. Nothing that seems enormous, thankfully, I'll see if I can come up with a fixed version we can test out, and upstream if it works fine.

The biggest issue is probably that I don't have an easy way to test on older versions of macOS, so will be taking a stab in them. I also don't know how far back rustc supports for the host (I know libstd supports back to 10.7, though). Also, does rustc fork itself during execution? If so, that may complicate things a little, at least in terms of testing.

Also, does rustc fork itself during execution?

No. It only spawns a new thread to run the compilation on.

fee1-dead · 2022-04-12T09:29:05Z

I don't know how I can move forward with this. Perhaps this could be left as an option? It uses more memory but is faster..

klensy · 2022-04-12T13:41:06Z

I don't know how I can move forward with this. Perhaps this could be left as an option? It uses more memory but is faster.

Rebase and bump mimalloc version? Maybe something new will be over that time.

Mark-Simulacrum · 2022-07-24T21:22:32Z

I think at this time we're unlikely to want to take a ~6-35% max-rss regression for the instruction count wins from adopting this PR.

That said, I'd love to see this rebased on a more recent version of mimalloc to see if they've managed to improve the memory overhead since then. If mimalloc is just making a different tradeoff on the memory/performance curve, then I'm not sure it's something we're willing to go for, given that rustc already uses a large amount of memory and we get some amount of complaints about that.

nnethercote · 2022-10-26T19:35:18Z

@fee1-dead: this was tested with libmimalloc-sys 0.1.23, which uses mimalloc 1.7.0. libmimalloc-sys 0.1.30 is now available, which uses mimalloc 1.7.6. That's not 2.0.x, and the release notes suggest that the difference between mimalloc 1.7.0 and 1.7.6 is mostly improved Mac support. Still, is it worth trying again?

@octavonce, @thomcc: Is there any prospect of libmimalloc-sys gaining support for 2.0.x?

thomcc · 2022-10-26T19:43:03Z

It already has support for it (As I note in purpleprotocol/mimalloc_rust#62 (comment), the mimalloc_rust submodule point towards 2.0.6 -- might be even newer now).

A big concern is that mimalloc's macOS override support is a lot less robust (and also a lot more dodgy) than jemalloc's. On other platforms it should be fine, though.

octavonce · 2022-10-26T19:56:59Z

@thomcc @nnethercote

A big concern is that mimalloc's macOS override support is a lot less robust (and also a lot more dodgy) than jemalloc's. On other platforms it should be fine, though.

But since this would be included as a static dependency to the compiler, and not overriding the OS allocator, I don't see how this is relevant in this case?

Maybe try to test again with the latest version?

thomcc · 2022-10-26T20:00:19Z

But since this would be included as a static dependency to the compiler, and not overriding the OS allocator, I don't see how this is relevant in this case?

That's not actually true, rustc needs to override malloc (as documented here and here). I've experimented with making it use #[global_allocator], but it's not exactly viable.

lqd · 2022-10-26T20:06:51Z

One of the 2.x betas was tried in #92317 but we should probably try to have up-to-date numbers.

octavonce · 2022-10-26T20:34:45Z

@thomcc Oh, alright. Thank you for the explanation, I thought it was straightforward enough

jq-rs · 2022-10-27T07:44:20Z

I think the API which allows using mimalloc in rustc, similar to jemalloc's one, is missing at the moment for the macOS part of mimalloc. Probably it should be implemented first. At least I am able to test the API if someone has a proto version available.

thomcc · 2022-10-27T08:17:34Z

I have a prototype version that fixes the issues with mimalloc's zone integration I can dig out. It's late now and I have work tomorrow, but I'll try to resurrect this weekend. Please ping me if I do not, or feel free to DM me on zulip if anybody wants more information on the issues with mimalloc's current malloc_zone implementation (TLDR the handling of zone free is wrong, and the interposing is highly dubious).

jq-rs · 2022-10-27T08:33:56Z

Thanks @thomcc. To elaborate a bit, I think some form of public _mi_macos_override_malloc is probably needed as a next step, it is now a static private function.

thomcc · 2022-10-27T14:25:00Z

I think it's implementation is at least as much of a problem than the privacy (it will behave very badly if any system APIs manage zones on their own, and I think we/LLVM are big enough that we probably need to worry about that).

rust-highfive assigned matthewjasper Dec 24, 2021

rustbot added T-compiler Relevant to the compiler team, which will review and decide on the PR/issue. T-rustdoc Relevant to the rustdoc team, which will review and decide on the PR/issue. labels Dec 24, 2021

rust-highfive added the S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. label Dec 24, 2021

This comment has been minimized.

Sign in to view

Replace jemalloc with mimalloc

b90cfc8

fee1-dead force-pushed the mimalloc branch from 663a864 to b90cfc8 Compare December 24, 2021 11:04

AngelicosPhosphoros reviewed Dec 24, 2021

View reviewed changes

fee1-dead marked this pull request as draft December 26, 2021 16:00

fee1-dead mentioned this pull request Dec 27, 2021

[Experiment] Replace jemalloc with mimalloc beta branch #92317

Closed

jq-rs reviewed Jan 5, 2022

View reviewed changes

Mark-Simulacrum assigned Mark-Simulacrum and unassigned matthewjasper Jul 24, 2022

Mark-Simulacrum added S-waiting-on-author Status: This is awaiting some action (such as code changes or more information) from the author. and removed S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. labels Jul 24, 2022

nnethercote mentioned this pull request Oct 26, 2022

Tracking issue for speeding up rustc via its build configuration #103595

Open

23 tasks

fee1-dead closed this Nov 21, 2022

fee1-dead deleted the mimalloc branch November 21, 2022 14:30

fee1-dead mentioned this pull request Nov 9, 2024

Use mimalloc in rustc #132796

Draft

Replace jemalloc with mimalloc #92249

Replace jemalloc with mimalloc #92249

Conversation

fee1-dead commented Dec 24, 2021 • edited Loading

rust-highfive commented Dec 24, 2021

This comment has been minimized.

dtolnay commented Dec 24, 2021

bors commented Dec 24, 2021

takkuumi commented Dec 24, 2021

AngelicosPhosphoros Dec 24, 2021 • edited Loading

Choose a reason for hiding this comment

AngelicosPhosphoros commented Dec 24, 2021

takkuumi commented Dec 24, 2021

fee1-dead commented Dec 24, 2021

AngelicosPhosphoros commented Dec 24, 2021 • edited Loading

bors commented Dec 24, 2021

takkuumi commented Dec 24, 2021

fee1-dead commented Dec 26, 2021

rust-timer commented Dec 26, 2021

rust-timer commented Dec 26, 2021

Mark-Simulacrum commented Dec 26, 2021

fee1-dead commented Dec 26, 2021

nikic commented Dec 26, 2021

fee1-dead commented Dec 26, 2021 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jq-rs Jan 9, 2022 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

thomcc Jan 9, 2022 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

fee1-dead commented Apr 12, 2022

klensy commented Apr 12, 2022

Mark-Simulacrum commented Jul 24, 2022

nnethercote commented Oct 26, 2022

thomcc commented Oct 26, 2022 • edited Loading

octavonce commented Oct 26, 2022 • edited Loading

thomcc commented Oct 26, 2022 • edited Loading

lqd commented Oct 26, 2022

octavonce commented Oct 26, 2022

jq-rs commented Oct 27, 2022

thomcc commented Oct 27, 2022 • edited Loading

jq-rs commented Oct 27, 2022

thomcc commented Oct 27, 2022

fee1-dead commented Dec 24, 2021 •

edited

Loading

AngelicosPhosphoros Dec 24, 2021 •

edited

Loading

AngelicosPhosphoros commented Dec 24, 2021 •

edited

Loading

fee1-dead commented Dec 26, 2021 •

edited

Loading

jq-rs Jan 9, 2022 •

edited

Loading

thomcc Jan 9, 2022 •

edited

Loading

thomcc commented Oct 26, 2022 •

edited

Loading

octavonce commented Oct 26, 2022 •

edited

Loading

thomcc commented Oct 26, 2022 •

edited

Loading

thomcc commented Oct 27, 2022 •

edited

Loading