| Summary: | Worklist should have a more fine grained way to just stop for some VMs | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Product: | WebKit | Reporter: | Saam Barati <saam> | ||||||
| Component: | JavaScriptCore | Assignee: | Saam Barati <saam> | ||||||
| Status: | NEW --- | ||||||||
| Severity: | Normal | CC: | benjamin, fpizlo, ggaren, gskachkov, guijemont, keith_miller, mark.lam, msaboff, rmorisset, ticaiolima, tzagallo, ysuzuki | ||||||
| Priority: | P2 | ||||||||
| Version: | WebKit Nightly Build | ||||||||
| Hardware: | Unspecified | ||||||||
| OS: | Unspecified | ||||||||
| Attachments: |
|
||||||||
|
Description
Saam Barati
2019-03-28 17:35:29 PDT
So I have a WIP patch. I'm testing on 2-core x86-64 MBP (using instruments to bring my core count down to 2 w/ hyper threading). The patch makes us go from total CPU time spent in "Heap::suspendCompilerThreads" from 34000 ms to 400ms. However, it's not a progression on bomb-workers. And there's a slight chance it's even a regression. I wonder if it's revealing other forms of contention that per-process locking was somehow saving us from. When I test in four core mode, it's also repeatably just a bit slower. However, we're also repeatably spending ~2 orders of magnitude less time in Heap::suspendCompilerThreads Created attachment 366233 [details]
WIP
Created attachment 366906 [details]
WIP
Stashing this WIP here.
|