Thanks Anton! Re: investigation of OOMs – the motivation for doing this work was based on bulk analysis of crash data coming into Edge from Windows devices. What I found was that a non-trivial percentage of renderer OOMs had the v8-oom-location crash key set to a value indicating we had exhausted the Oilpan heap, and of those crashes a large fraction showed plenty of available commit on the system. I can certainly believe that some fraction of these crashes are leaks. I can sample a few dumps to get a better sense, but I’m not sure how to distinguish a leak in a post-mortem context – any suggestions on what to look for?
Re: performance – I ran a few Pinpoint jobs on a change implementing the 3-bit shift plus a larger reservation to ensure the correct bit pattern for a 4GB cage. Speedometer was perf-neutral. blink_perf.bindings and blink_perf.css showed a mix of improvements and regressions. Based on my past experience with the blink_perf.* benchmarks I’d say the results are within the margin of error, but let me know if you see something concerning or would like me to run some additional tests. Speedometer: https://pinpoint-dot-chromeperf.appspot.com/job/13c2308e460000 blink_perf.bindings: https://pinpoint-dot-chromeperf.appspot.com/job/15b7a7ba460000 blink_perf.css: https://pinpoint-dot-chromeperf.appspot.com/job/16b8dc08460000 Kevin From: v8-dev@googlegroups.com<mailto:v8-dev@googlegroups.com> <v8-dev@googlegroups.com<mailto:v8-dev@googlegroups.com>> On Behalf Of Anton Bikineev Sent: Thursday, March 30, 2023 8:11 AM To: v8-dev <v8-dev@googlegroups.com<mailto:v8-dev@googlegroups.com>> Subject: [v8-dev] Re: Design doc: Oilpan Caged Heap Expansion (looks like I sent the message privately, repeating here for visibility) Thanks Kevin for writing up the proposal! I have a couple of comments/concerns about the proposal. Have you investigated the OOMs resulting from Oilpan cage exhaustion? Since we returned the 4GB cage, we rarely see any in Chrome. Such OOMs usually indicate leaks. Small cages actually help manifest and fix them earlier. Re the shift and the 100KB increase. The branchless design and the alignment bits do seamlessly allow to increase the cage size. However, there may be a performance penalty associated: * Since the decompression snippet becomes larger, this may add additional pressure on instruction cache. * CyclesPerInstruction of add is known<https://www.agner.org/optimize/instruction_tables.pdf> to be 2x smaller than that of shl/sal, even on the modern IceLake/TigerLake. This may have negative impact on the workloads that have non-dependent decompressions, e.g.: for (auto* member: heap_vector) member->do_something(); I assume this is why llvm performs<https://github.com/llvm/llvm-project/blob/main/llvm/lib/Target/X86/X86InstrCompiler.td#L1842> this optimization for the 4GB cage. On Thursday, March 30, 2023 at 1:33:08 AM UTC+2 kbab...@microsoft.com<mailto:kbab...@microsoft.com> wrote: Hi all, I wrote a design doc with a proposal to give embedders an option to expand the Oilpan caged heap: https://docs.google.com/document/d/1yGAsu_41rU8_hGQ9tcSKH84Em3vj3uzw_c0YlL7SCjA/edit?usp=sharing Thanks, Kevin -- -- v8-dev mailing list v8-dev@googlegroups.com<mailto:v8-dev@googlegroups.com> http://groups.google.com/group/v8-dev --- You received this message because you are subscribed to the Google Groups "v8-dev" group. To unsubscribe from this group and stop receiving emails from it, send an email to v8-dev+unsubscr...@googlegroups.com<mailto:v8-dev+unsubscr...@googlegroups.com>. To view this discussion on the web visit https://groups.google.com/d/msgid/v8-dev/0abb08d9-b724-4289-bebb-88125ef05d11n%40googlegroups.com<https://groups.google.com/d/msgid/v8-dev/0abb08d9-b724-4289-bebb-88125ef05d11n%40googlegroups.com?utm_medium=email&utm_source=footer>. -- -- v8-dev mailing list v8-dev@googlegroups.com http://groups.google.com/group/v8-dev --- You received this message because you are subscribed to the Google Groups "v8-dev" group. To unsubscribe from this group and stop receiving emails from it, send an email to v8-dev+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/v8-dev/PH0PR00MB132954E3671C545A2F631494C0909%40PH0PR00MB1329.namprd00.prod.outlook.com.
