1940320 - Too many FULL_WASM_ANYREF_BUFFER minor GC

Status: RESOLVED FIXED

(Core :: JavaScript: WebAssembly, enhancement, P1)

Description

[Yury Delendik (:yury)]

Attachment: JS_GC_REPORT_PRETENURE.txt.gz

When running dart's benchmark (dart-wasm-gc-benchmarks), the SpiderMonkey generates too many minor GCs with reason of FULL_WASM_ANYREF_BUFFER. It looks like a legitimate behavior. The minor GC and object promotion operations are top calls in the profiler https://share.firefox.dev/4gHWMMF

Increasing AnyRef store buffer did not improve the benchmark test scores.

Per conversation with sfink, changing highPromotionRate to 0.4 at https://searchfox.org/mozilla-central/source/js/src/gc/Pretenuring.cpp#325 improved the peformance by 23% (was average frame time 8ms, and after the change 6.5ms). The amount of FULL_WASM_ANYREF_BUFFER decreased 11x. Additionally, applying 0.3 or 0.5 did not improve the situation much.

Do we want to change the highPromotionRate for Wasm GC objects?

Attaching log with JS_GC_REPORT_PRETENURE=1, which has some promotional rate numbers for benchmark run without any changes applied.

Comment 1

[Jon Coppeard (:jonco)]

(In reply to Yury Delendik (:yury) from comment #0)
Currently the benchmark spends 22% of its time collecting the nursery, which does seem high.

Per conversation with sfink, changing highPromotionRate to 0.4 at https://searchfox.org/mozilla-central/source/js/src/gc/Pretenuring.cpp#325 improved the peformance by 23%

Changing this may well make other benchmarks slower. And the fact that you would need to pick a very specific value to improve this one indicates that this is not a great approach.

Some other possibilities that may improve this:

• more finegrained tracking of allocation sites for Wasm GC objects
• some analysis in Wasm to reduce nursery allocation of objects that will always end up tenured
• changes to the nursery to reduce the cost of allocating long lived objects there (e.g. partial nursery eviction)

Comment 2

[Steve Fink [:sfink] [:s:]]

To be clear, it's not that 40% is the magic number. 50% had some improvement, just not as much as 40%. 30% and 40% behaved the same. So the improvement requires 40% or less, not exactly 40%.

Also, here's JS_GC_REPORT_PRETENURE output from yury:

Pretenuring info after minor GC 20 for FULL_WASM_ANYREF_BUFFER reason with promotion rate 15.4%:
  Site             Zone             Location             BytecodeOp   SiteKind  TraceKind NAllocs  Promotes PRate  State     
       0x1041a9930      0x102b8d600                                   normal    JS Object      309        0   0.0% ShortLived
       0x1041a9980      0x102b8d600                                   normal    JS Object      309        0   0.0% ShortLived
       0x10419c140      0x102b8d600                                   normal    JS Object      254        0   0.0% ShortLived
       0x1041cc3e0      0x102b8d600                                   normal    JS Object      294        0   0.0% ShortLived
       0x1041d0d00      0x102b8d600                                   normal    JS Object      636        0   0.0% ShortLived
       0x1041c26b0      0x102b8d600                                   normal    JS Object      652        0   0.0% ShortLived
       0x1041b09b0      0x102b8d600                                   normal    JS Object      255      136  53.3% ShortLived
       0x1041b0690      0x102b8d600                                   normal    JS Object      255      136  53.3% ShortLived
       0x1041cde70      0x102b8d600                                   normal    JS Object      302      128  42.4% ShortLived
       0x1041cd510      0x102b8d600                                   normal    JS Object      302      128  42.4% ShortLived
       0x1041c3e70      0x102b8d600                                   normal    JS Object      388        0   0.0% ShortLived
       0x1041b6860      0x102b8d600                                   normal    JS Object      430        0   0.0% ShortLived
       0x1041af1a0      0x102b8d600                                   normal    JS Object      320       41  12.8% ShortLived
       0x1041ad3f0      0x102b8d600                                   normal    JS Object      252        0   0.0% ShortLived
       0x1041944e0      0x102b8d600                                   normal    JS Object      288        0   0.0% ShortLived
       0x1041949e0      0x102b8d600                                   normal    JS Object      457        0   0.0% ShortLived
       0x1041cd4c0      0x102b8d600                                   normal    JS Object     1003      248  24.7% ShortLived
       0x1041cd740      0x102b8d600                                   normal    JS Object     1003      248  24.7% ShortLived
       0x1041cdce0      0x102b8d600                                   normal    JS Object      819      136  16.6% ShortLived
       0x104194a30      0x102b8d600                                   normal    JS Object      457        0   0.0% ShortLived
       0x104192460      0x102b8d600                                   normal    JS Object      336        0   0.0% ShortLived
       0x104192320      0x102b8d600                                   normal    JS Object      384        0   0.0% ShortLived
       0x1041950c0      0x102b8d600                                   normal    JS Object      932        0   0.0% ShortLived
       0x1041c55e0      0x102b8d600                                   normal    JS Object     3410        0   0.0% ShortLived
       0x1041ccde0      0x102b8d600                                   normal    JS Object     1232        0   0.0% ShortLived
       0x1041ccf70      0x102b8d600                                   normal    JS Object     1232        0   0.0% ShortLived
       0x1041c5540      0x102b8d600                                   normal    JS Object     3443        0   0.0% ShortLived
       0x1041ce000      0x102b8d600                                   normal    JS Object     1894      304  16.1% ShortLived
       0x1041ce370      0x102b8d600                                   normal    JS Object     2220      304  13.7% ShortLived
       0x1041cd6f0      0x102b8d600                                   normal    JS Object     2220      304  13.7% ShortLived
       0x104192370      0x102b8d600                                   normal    JS Object      384        0   0.0% ShortLived
       0x1041924b0      0x102b8d600                                   normal    JS Object      384        0   0.0% ShortLived
       0x1041b4830      0x102b8d600                                   normal    JS Object      390        0   0.0% ShortLived
       0x1041ae340      0x102b8d600                                   normal    JS Object      234      117  50.0% ShortLived
       0x104191830      0x102b8d600                                   normal    JS Object      315        0   0.0% ShortLived
       0x1041a5560      0x102b8d600                                   normal    JS Object      390        0   0.0% ShortLived
       0x1041a55b0      0x102b8d600                                   normal    JS Object      390        0   0.0% ShortLived
       0x1041d0760      0x102b8d600                                   normal    JS Object      349        0   0.0% ShortLived
       0x1041cc2a0      0x102b8d600                                   normal    JS Object     4352        1   0.0% ShortLived
       0x1041bf460      0x102b8d600                                   normal    JS Object     1414        1   0.1% ShortLived
       0x1041d15c0      0x102b8d600                                   normal    JS Object     2546      394  15.5% ShortLived
       0x1041d02b0      0x102b8d600                                   normal    JS Object     7514     2638  35.1% ShortLived
       0x1041d1520      0x102b8d600                                   normal    JS Object     4048     1584  39.1% ShortLived
       0x104190160      0x102b8d600                                   normal    JS Object     5169     1848  35.8% ShortLived
       0x1041abc30      0x102b8d600                                   normal    JS Object      273       39  14.3% ShortLived
       0x1041abb90      0x102b8d600                                   normal    JS Object      273       39  14.3% ShortLived
       0x1041d0f80      0x102b8d600                                   normal    JS Object    11850     2720  23.0% ShortLived
       0x1041d14d0      0x102b8d600                                   normal    JS Object    10752     1127  10.5% ShortLived
       0x1041cdab0      0x102b8d600                                   normal    JS Object      324        0   0.0% ShortLived
       0x104191e20      0x102b8d600                                   normal    JS Object      324       16   4.9% ShortLived
       0x1041d12a0      0x102b8d600                                   normal    JS Object    11299     5663  50.1% ShortLived
       0x1041d1390      0x102b8d600                                   normal    JS Object    25296      394   1.6% ShortLived
       0x1041d1340      0x102b8d600                                   normal    JS Object    24046      757   3.1% ShortLived
       0x102b8de08      0x102b8d600                                   unknown   JS Object        4        5        Unknown  

Comment 3

[Yury Delendik (:yury)]

After instrumentation of the spidermonkey, the benchmark indeed has some types/allocSites with multiple callsites. About 10-20% of types in the dart benchmark. It is worth to try assigning individual allocSite to a callsite/bytecode.

Comment 4

[Yury Delendik (:yury)]

Attachment: Bug 1940320 - Add AllocSite for every new Wasm GC op. r?rhunt

Comment 5

[Yury Delendik (:yury)]

Some numbers for WIP runs:

AverageDraw	Before	After	Change
baseline	11074 us	10401 us	-7%
ion	8074 us	7484 us	-7%

The amount of minor GCs dropped from 1083 -> 580, 50%.

There is some issue in tiering mode I need to investigate, and further optimize the WIP.

Comment 6

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/5d80c7f6af23
Add AllocSite for every new Wasm GC op. r=rhunt

Comment 7

[Cosmin Sabou [:CosminS]]

https://hg.mozilla.org/mozilla-central/rev/5d80c7f6af23