1978645 - Consider using BufferAllocator for OOL data for wasm GC objects

Status: RESOLVED FIXED

(Core :: JavaScript: WebAssembly, enhancement, P3)

Description

[Ryan Hunt [:rhunt]]

The GC now has support for a BufferAllocator [1] which can be used to allocate variable sized extra data and associate it with a GC object. This looks like it could replace our use of the Malloc'edBlockCache [2]. We should see how well it performs if we were to switch.

[1] https://searchfox.org/mozilla-central/rev/387160feb07b75ae76bfc12df035a10f58d25168/js/src/gc/BufferAllocator.h#195
[2] https://searchfox.org/mozilla-central/rev/387160feb07b75ae76bfc12df035a10f58d25168/js/src/gc/MallocedBlockCache.h#19

Comment 1

[Jon Coppeard (:jonco)]

Attachment: Bug 1978645 - Part 1: Use buffer allocator for wasm struct outline data r?jseward

Comment 2

[Jon Coppeard (:jonco)]

Attachment: Bug 1978645 - Part 2: Use buffer allocator for wasm array data r?jseward

Comment 3

[Jon Coppeard (:jonco)]

Attachment: Bug 1978645 - Part 3: Remove malloc block cache and nursery trailer tracking r?jseward

Comment 4

[Jon Coppeard (:jonco)]

I ran some benchmarks on these patches in the shell, testing on x64 Linux and arm64 macOS. This used nightly builds where the poisoning behaviour is the same for both jemalloc and the buffer allocator.

                           Linux                  macOS
                           Score  Major  Minor    Score  Major  Minor
                                  time   time            time   time

  Dart-flute-complex-wasm  +5.5%  -13%   -18%     +6.6%  -8.4%  -28%     
  Dart-flute-todomvc-wasm  +3.1%  +5%    -28%     +3.7%  +50%   -32%
  Kotlin-compose-wasm      +3.2%  -      -2.8%    -1%    +6.6%  -12%         

I did some profiling on a browser build (macOS) to check nothing untoward was happening. These show a significant reduction in main thread minor GC time due to not sweeping trailer blocks there and reduction in helper thread time calling object finalizers, with a smaller increase in helper thread time for buffer allocator sweeping.

I don't know why Dart-flute-todomvc-wasm shows increased major GC time in the shell. This didn't happen in browser testing.

Comment 5

[Jon Coppeard (:jonco)]

I had to limit the maximum buffer size to allocate directly in the nursery because j2cl-box2d-wasm was filling the nursery with these leading to longer nursery collections which lead to reducing the nursery size to compensate. Updated results with this change:

                           macOS             Linux
                           Score   Total GC  Score   Total GC   
                                   time              time   

  Dart-flute-complex-wasm  +6.7%   -24%      +5.3%   -17%
  Dart-flute-todomvc-wasm  +3.1%   -24%      +3.3%   -22%
  Kotlin-compose-wasm      +3.1%   -5.8%     +4.4%   -5.6%
  j2cl-box2d-wasm          -       -12%      -       -18%

Comment 6

[Jon Coppeard (:jonco)]

Attachment: Bug 1978645 - Part 0: Add argument for max nursery size to buffer allocation APIs r?sfink

Wasm GC allocations tend to get tenured more often than regular JS objects
which means minor GCs can spend a lot of time moving these directly allocated
buffers. This increases minor GC time which also means we keep the nursery size
smaller, affecting performance. This adds an option to set the maximum size of
buffer allocations made directly in the nursery. For Wasm GC allocations we'll
make this smaller than the regular size.

Comment 7

[Pulsebot]

Pushed by [e-mail]:
https://github.com/mozilla-firefox/firefox/commit/c5c8110cd319
https://hg.mozilla.org/integration/autoland/rev/9be0b8b761ee
Part 0: Add argument for max nursery size to buffer allocation APIs r=sfink
https://github.com/mozilla-firefox/firefox/commit/4c2e113134d4
https://hg.mozilla.org/integration/autoland/rev/19e4500ea7f6
Part 1: Use buffer allocator for wasm struct outline data r=jseward
https://github.com/mozilla-firefox/firefox/commit/1884530df58f
https://hg.mozilla.org/integration/autoland/rev/0a7420a74a64
Part 2: Use buffer allocator for wasm array data r=jseward
https://github.com/mozilla-firefox/firefox/commit/a39d39fa905d
https://hg.mozilla.org/integration/autoland/rev/9aaf56b12eab
Part 3: Remove malloc block cache and nursery trailer tracking r=jseward

Comment 8

[Norisz Fay [:noriszfay]]

https://hg.mozilla.org/mozilla-central/rev/9be0b8b761ee
https://hg.mozilla.org/mozilla-central/rev/19e4500ea7f6
https://hg.mozilla.org/mozilla-central/rev/0a7420a74a64
https://hg.mozilla.org/mozilla-central/rev/9aaf56b12eab