227001 2021-06-14 16:56:33 -0700 Too much CPU time is spent under MemoryPressureHandler::currentMemoryUsagePolicy() 2021-06-21 14:38:07 -0700 1 1 1 Unclassified WebKit Layout and Rendering WebKit Nightly Build Unspecified Unspecified RESOLVED FIXED InRadar P2 Normal --- 1 cdumez cdumez bfulgham changseok esprehn+autocc ews-watchlist fred.wang ggaren glenn kondapallykalyan nham pdr simon.fraser webkit-bug-importer zalan oldest_to_newest 1769626 0 cdumez 2021-06-14 16:56:33 -0700 Too much CPU time is spent under MemoryPressureHandler::currentMemoryUsagePolicy(). Looking at profiles of our WebContent processes, we frequently see traces like these: Sample Count, Samples %, Normalized CPU %, Symbol 31, 0.0%, 0.0%, WebCore::RenderLayerCompositor::cacheAcceleratedCompositingFlags() (in WebCore) 30, 0.0%, 0.0%, WTF::MemoryPressureHandler::currentMemoryUsagePolicy() (in JavaScriptCore) 30, 0.0%, 0.0%, task_info (in libsystem_kernel.dylib) It is a decent amount of samples and shows several times per process. We should try and minimize calls to WTF::MemoryPressureHandler::currentMemoryUsagePolicy() to reduce CPU usage. 1769689 1 nham 2021-06-14 21:48:48 -0700 Maybe we should just ditch computing the memory footprint on every call and map MemoryUsagePolicy to OS memory pressure levels (e.g. unrestricted => normal pressure, conservative => warning pressure, strict => critical pressure). Then we could just rely on the OS memory pressure notification handler to update MemoryUsagePolicy. I guess the other obvious thing to do here is to cache the memoryFootprint for X seconds. 1771434 2 431904 cdumez 2021-06-21 13:55:08 -0700 Created attachment 431904 Patch 1771438 3 431904 simon.fraser 2021-06-21 14:03:22 -0700 Comment on attachment 431904 Patch 10s seems too long. Maybe 1s or 2s? 1771439 4 simon.fraser 2021-06-21 14:03:46 -0700 Or we could do it once per rendering update. 1771440 5 431906 cdumez 2021-06-21 14:06:02 -0700 Created attachment 431906 Patch 1771443 6 ggaren 2021-06-21 14:08:23 -0700 Can we pick the delay based on what it would take to make the % time in a representative profile negligible? In the snippet posted here, the % time is already 0.0% (rounded, I assume), so I'm not sure any delay would be justified. To say out loud what is perhaps implicit so far: memory pressure can change rapidly, so there is some benefit to responding as quickly as possible (but not so quickly that we waste CPU). 1771445 7 simon.fraser 2021-06-21 14:09:24 -0700 (In reply to Geoffrey Garen from comment #6) > Can we pick the delay based on what it would take to make the % time in a > representative profile negligible? > > In the snippet posted here, the % time is already 0.0% (rounded, I assume), > so I'm not sure any delay would be justified. > > To say out loud what is perhaps implicit so far: memory pressure can change > rapidly, so there is some benefit to responding as quickly as possible (but > not so quickly that we waste CPU). I suspect that the only cases where it's non-negligible are those where there are many forced layouts per rendering update. So once per rendering update feels right to me. 1771446 8 cdumez 2021-06-21 14:09:32 -0700 (In reply to Geoffrey Garen from comment #6) > Can we pick the delay based on what it would take to make the % time in a > representative profile negligible? > > In the snippet posted here, the % time is already 0.0% (rounded, I assume), > so I'm not sure any delay would be justified. This shows many times per profile. Not sure how to aggregate all the calls in the profiler. > > To say out loud what is perhaps implicit so far: memory pressure can change > rapidly, so there is some benefit to responding as quickly as possible (but > not so quickly that we waste CPU). 1771447 9 cdumez 2021-06-21 14:12:24 -0700 (In reply to Chris Dumez from comment #8) > (In reply to Geoffrey Garen from comment #6) > > Can we pick the delay based on what it would take to make the % time in a > > representative profile negligible? > > > > In the snippet posted here, the % time is already 0.0% (rounded, I assume), > > so I'm not sure any delay would be justified. > > This shows many times per profile. Not sure how to aggregate all the calls > in the profiler. > > > > > To say out loud what is perhaps implicit so far: memory pressure can change > > rapidly, so there is some benefit to responding as quickly as possible (but > > not so quickly that we waste CPU). If the point is to respond to memory pressure, then we have the memory pressure warning/signal for that. This code relies on a certain % of the memory being used, which is not the same as memory pressure. 1771450 10 cdumez 2021-06-21 14:15:45 -0700 (In reply to Chris Dumez from comment #9) > (In reply to Chris Dumez from comment #8) > > (In reply to Geoffrey Garen from comment #6) > > > Can we pick the delay based on what it would take to make the % time in a > > > representative profile negligible? > > > > > > In the snippet posted here, the % time is already 0.0% (rounded, I assume), > > > so I'm not sure any delay would be justified. > > > > This shows many times per profile. Not sure how to aggregate all the calls > > in the profiler. > > > > > > > > To say out loud what is perhaps implicit so far: memory pressure can change > > > rapidly, so there is some benefit to responding as quickly as possible (but > > > not so quickly that we waste CPU). > > If the point is to respond to memory pressure, then we have the memory > pressure warning/signal for that. This code relies on a certain % of the > memory being used, which is not the same as memory pressure. And this threshold appears to be 33% of RAM size of macOS and 50% of RAM size on iOS. 1771458 11 ews-feeder 2021-06-21 14:36:34 -0700 Committed r279084 (239003@main): <https://commits.webkit.org/239003@main> All reviewed patches have been landed. Closing bug and clearing flags on attachment 431906. 1771459 12 webkit-bug-importer 2021-06-21 14:38:07 -0700 <rdar://problem/79581141> 431904 2021-06-21 13:55:08 -0700 2021-06-21 14:06:00 -0700 Patch bug-227001-20210621135507.patch text/plain 2334 cdumez U3VidmVyc2lvbiBSZXZpc2lvbjogMjc5MDcwCmRpZmYgLS1naXQgYS9Tb3VyY2UvV2ViQ29yZS9D aGFuZ2VMb2cgYi9Tb3VyY2UvV2ViQ29yZS9DaGFuZ2VMb2cKaW5kZXggNzYzMzM2NmNmNGZlZDUw YzBlZWZhYzhjOGQ5ZDE4ZTk4OTA2ZDRmMi4uZjMyMWYwNzBjNGFmZDM2OWU0NzQwNzMwMTI4MTAz ZTIzOGE1NGI0NSAxMDA2NDQKLS0tIGEvU291cmNlL1dlYkNvcmUvQ2hhbmdlTG9nCisrKyBiL1Nv dXJjZS9XZWJDb3JlL0NoYW5nZUxvZwpAQCAtMSwzICsxLDE3IEBACisyMDIxLTA2LTIxICBDaHJp cyBEdW1leiAgPGNkdW1lekBhcHBsZS5jb20+CisKKyAgICAgICAgVG9vIG11Y2ggQ1BVIHRpbWUg aXMgc3BlbnQgdW5kZXIgTWVtb3J5UHJlc3N1cmVIYW5kbGVyOjpjdXJyZW50TWVtb3J5VXNhZ2VQ b2xpY3koKQorICAgICAgICBodHRwczovL2J1Z3Mud2Via2l0Lm9yZy9zaG93X2J1Zy5jZ2k/aWQ9 MjI3MDAxCisKKyAgICAgICAgUmV2aWV3ZWQgYnkgTk9CT0RZIChPT1BTISkuCisKKyAgICAgICAg TWFrZSBpdCBzbyB0aGF0IFJlbmRlckxheWVyQ29tcG9zaXRvcjo6dXBkYXRlQ29tcG9zaXRpbmdQ b2xpY3koKSBxdWVyaWVzIE1lbW9yeVByZXNzdXJlSGFuZGxlcjo6Y3VycmVudE1lbW9yeVVzYWdl UG9saWN5KCkKKyAgICAgICAgYXQgbW9zdCBvbmNlIGV2ZXJ5IDEwIHNlY29uZHMgZ2l2ZW4gdGhh dCB0aGUgY2FsbCBpcyBleHBlbnNpdmUuIEl0IGNhbiBjdXJyZW50bHkgZ2V0IGNhbGxlZCBzZXZl cmFsIHRpbWVzIGEgc2Vjb25kIGFuZAorICAgICAgICBzaG93cyBvbiBDUFUgcHJvZmlsZXMuCisK KyAgICAgICAgKiByZW5kZXJpbmcvUmVuZGVyTGF5ZXJDb21wb3NpdG9yLmNwcDoKKyAgICAgICAg KFdlYkNvcmU6OlJlbmRlckxheWVyQ29tcG9zaXRvcjo6dXBkYXRlQ29tcG9zaXRpbmdQb2xpY3kp OgorCiAyMDIxLTA2LTIxICBSb2IgQnVpcyAgPHJidWlzQGlnYWxpYS5jb20+CiAKICAgICAgICAg Q1NTT00gdGVzdCBmb3Igc2VyaWFsaXppbmcgZm9udC12YXJpYW50IGZhaWxzCmRpZmYgLS1naXQg YS9Tb3VyY2UvV2ViQ29yZS9yZW5kZXJpbmcvUmVuZGVyTGF5ZXJDb21wb3NpdG9yLmNwcCBiL1Nv dXJjZS9XZWJDb3JlL3JlbmRlcmluZy9SZW5kZXJMYXllckNvbXBvc2l0b3IuY3BwCmluZGV4IDEx YzlmZjdhZDRlYjNmZTFlMmIxMTg3MGQ5NDI4NTI4ZGZlZTU1ZTYuLmZlM2E3Y2ZmZTI5NDBiNmZk NTU4NWNhN2VjYjRiNTAxZDU0YjZjZjkgMTAwNjQ0Ci0tLSBhL1NvdXJjZS9XZWJDb3JlL3JlbmRl cmluZy9SZW5kZXJMYXllckNvbXBvc2l0b3IuY3BwCisrKyBiL1NvdXJjZS9XZWJDb3JlL3JlbmRl cmluZy9SZW5kZXJMYXllckNvbXBvc2l0b3IuY3BwCkBAIC00ODUsOCArNDg1LDE2IEBAIGJvb2wg UmVuZGVyTGF5ZXJDb21wb3NpdG9yOjp1cGRhdGVDb21wb3NpdGluZ1BvbGljeSgpCiAgICAgICAg IHJldHVybiBtX2NvbXBvc2l0aW5nUG9saWN5ICE9IGN1cnJlbnRQb2xpY3k7CiAgICAgfQogICAg IAotICAgIGF1dG8gbWVtb3J5UG9saWN5ID0gTWVtb3J5UHJlc3N1cmVIYW5kbGVyOjpjdXJyZW50 TWVtb3J5VXNhZ2VQb2xpY3koKTsKLSAgICBtX2NvbXBvc2l0aW5nUG9saWN5ID0gbWVtb3J5UG9s aWN5ID09IFdURjo6TWVtb3J5VXNhZ2VQb2xpY3k6OlVucmVzdHJpY3RlZCA/IENvbXBvc2l0aW5n UG9saWN5OjpOb3JtYWwgOiBDb21wb3NpdGluZ1BvbGljeTo6Q29uc2VydmF0aXZlOworICAgIHN0 YXRpYyBhdXRvIGNhY2hlZE1lbW9yeVBvbGljeSA9IFdURjo6TWVtb3J5VXNhZ2VQb2xpY3k6OlVu cmVzdHJpY3RlZDsKKyAgICBzdGF0aWMgTW9ub3RvbmljVGltZSBjYWNoZWRNZW1vcnlQb2xpY3lU aW1lOworICAgIHN0YXRpYyBjb25zdGV4cHIgYXV0byBtZW1vcnlQb2xpY3lDYWNoaW5nRHVyYXRp b24gPSAxMF9zOworICAgIGF1dG8gbm93ID0gTW9ub3RvbmljVGltZTo6bm93KCk7CisgICAgaWYg KG5vdyAtIGNhY2hlZE1lbW9yeVBvbGljeVRpbWUgPiBtZW1vcnlQb2xpY3lDYWNoaW5nRHVyYXRp b24pIHsKKyAgICAgICAgY2FjaGVkTWVtb3J5UG9saWN5ID0gTWVtb3J5UHJlc3N1cmVIYW5kbGVy OjpjdXJyZW50TWVtb3J5VXNhZ2VQb2xpY3koKTsKKyAgICAgICAgY2FjaGVkTWVtb3J5UG9saWN5 VGltZSA9IG5vdzsKKyAgICB9CisKKyAgICBtX2NvbXBvc2l0aW5nUG9saWN5ID0gY2FjaGVkTWVt b3J5UG9saWN5ID09IFdURjo6TWVtb3J5VXNhZ2VQb2xpY3k6OlVucmVzdHJpY3RlZCA/IENvbXBv c2l0aW5nUG9saWN5OjpOb3JtYWwgOiBDb21wb3NpdGluZ1BvbGljeTo6Q29uc2VydmF0aXZlOwog ICAgIHJldHVybiBtX2NvbXBvc2l0aW5nUG9saWN5ICE9IGN1cnJlbnRQb2xpY3k7CiB9CiAK 431906 2021-06-21 14:06:02 -0700 2021-06-21 14:36:36 -0700 Patch bug-227001-20210621140601.patch text/plain 2330 cdumez U3VidmVyc2lvbiBSZXZpc2lvbjogMjc5MDcwCmRpZmYgLS1naXQgYS9Tb3VyY2UvV2ViQ29yZS9D aGFuZ2VMb2cgYi9Tb3VyY2UvV2ViQ29yZS9DaGFuZ2VMb2cKaW5kZXggNzYzMzM2NmNmNGZlZDUw YzBlZWZhYzhjOGQ5ZDE4ZTk4OTA2ZDRmMi4uMDhhNjRlOWJhNWRmODA5OWUwNDJmOTcwZTllZDFh OTUyOGYxODAwZSAxMDA2NDQKLS0tIGEvU291cmNlL1dlYkNvcmUvQ2hhbmdlTG9nCisrKyBiL1Nv dXJjZS9XZWJDb3JlL0NoYW5nZUxvZwpAQCAtMSwzICsxLDE3IEBACisyMDIxLTA2LTIxICBDaHJp cyBEdW1leiAgPGNkdW1lekBhcHBsZS5jb20+CisKKyAgICAgICAgVG9vIG11Y2ggQ1BVIHRpbWUg aXMgc3BlbnQgdW5kZXIgTWVtb3J5UHJlc3N1cmVIYW5kbGVyOjpjdXJyZW50TWVtb3J5VXNhZ2VQ b2xpY3koKQorICAgICAgICBodHRwczovL2J1Z3Mud2Via2l0Lm9yZy9zaG93X2J1Zy5jZ2k/aWQ9 MjI3MDAxCisKKyAgICAgICAgUmV2aWV3ZWQgYnkgU2ltb24gRnJhc2VyLgorCisgICAgICAgIE1h a2UgaXQgc28gdGhhdCBSZW5kZXJMYXllckNvbXBvc2l0b3I6OnVwZGF0ZUNvbXBvc2l0aW5nUG9s aWN5KCkgcXVlcmllcyBNZW1vcnlQcmVzc3VyZUhhbmRsZXI6OmN1cnJlbnRNZW1vcnlVc2FnZVBv bGljeSgpCisgICAgICAgIGF0IG1vc3Qgb25jZSBldmVyeSAyIHNlY29uZHMgZ2l2ZW4gdGhhdCB0 aGUgY2FsbCBpcyBleHBlbnNpdmUuIEl0IGNhbiBjdXJyZW50bHkgZ2V0IGNhbGxlZCBzZXZlcmFs IHRpbWVzIGEgc2Vjb25kIGFuZAorICAgICAgICBzaG93cyBvbiBDUFUgcHJvZmlsZXMuCisKKyAg ICAgICAgKiByZW5kZXJpbmcvUmVuZGVyTGF5ZXJDb21wb3NpdG9yLmNwcDoKKyAgICAgICAgKFdl YkNvcmU6OlJlbmRlckxheWVyQ29tcG9zaXRvcjo6dXBkYXRlQ29tcG9zaXRpbmdQb2xpY3kpOgor CiAyMDIxLTA2LTIxICBSb2IgQnVpcyAgPHJidWlzQGlnYWxpYS5jb20+CiAKICAgICAgICAgQ1NT T00gdGVzdCBmb3Igc2VyaWFsaXppbmcgZm9udC12YXJpYW50IGZhaWxzCmRpZmYgLS1naXQgYS9T b3VyY2UvV2ViQ29yZS9yZW5kZXJpbmcvUmVuZGVyTGF5ZXJDb21wb3NpdG9yLmNwcCBiL1NvdXJj ZS9XZWJDb3JlL3JlbmRlcmluZy9SZW5kZXJMYXllckNvbXBvc2l0b3IuY3BwCmluZGV4IDExYzlm ZjdhZDRlYjNmZTFlMmIxMTg3MGQ5NDI4NTI4ZGZlZTU1ZTYuLmViZWI1ZTE2YmNkMTY2OWZiZTJm MWM1YWE5ZDM1ODE4NzQzMjIwNDkgMTAwNjQ0Ci0tLSBhL1NvdXJjZS9XZWJDb3JlL3JlbmRlcmlu Zy9SZW5kZXJMYXllckNvbXBvc2l0b3IuY3BwCisrKyBiL1NvdXJjZS9XZWJDb3JlL3JlbmRlcmlu Zy9SZW5kZXJMYXllckNvbXBvc2l0b3IuY3BwCkBAIC00ODUsOCArNDg1LDE2IEBAIGJvb2wgUmVu ZGVyTGF5ZXJDb21wb3NpdG9yOjp1cGRhdGVDb21wb3NpdGluZ1BvbGljeSgpCiAgICAgICAgIHJl dHVybiBtX2NvbXBvc2l0aW5nUG9saWN5ICE9IGN1cnJlbnRQb2xpY3k7CiAgICAgfQogICAgIAot ICAgIGF1dG8gbWVtb3J5UG9saWN5ID0gTWVtb3J5UHJlc3N1cmVIYW5kbGVyOjpjdXJyZW50TWVt b3J5VXNhZ2VQb2xpY3koKTsKLSAgICBtX2NvbXBvc2l0aW5nUG9saWN5ID0gbWVtb3J5UG9saWN5 ID09IFdURjo6TWVtb3J5VXNhZ2VQb2xpY3k6OlVucmVzdHJpY3RlZCA/IENvbXBvc2l0aW5nUG9s aWN5OjpOb3JtYWwgOiBDb21wb3NpdGluZ1BvbGljeTo6Q29uc2VydmF0aXZlOworICAgIHN0YXRp YyBhdXRvIGNhY2hlZE1lbW9yeVBvbGljeSA9IFdURjo6TWVtb3J5VXNhZ2VQb2xpY3k6OlVucmVz dHJpY3RlZDsKKyAgICBzdGF0aWMgTW9ub3RvbmljVGltZSBjYWNoZWRNZW1vcnlQb2xpY3lUaW1l OworICAgIHN0YXRpYyBjb25zdGV4cHIgYXV0byBtZW1vcnlQb2xpY3lDYWNoaW5nRHVyYXRpb24g PSAyX3M7CisgICAgYXV0byBub3cgPSBNb25vdG9uaWNUaW1lOjpub3coKTsKKyAgICBpZiAobm93 IC0gY2FjaGVkTWVtb3J5UG9saWN5VGltZSA+IG1lbW9yeVBvbGljeUNhY2hpbmdEdXJhdGlvbikg eworICAgICAgICBjYWNoZWRNZW1vcnlQb2xpY3kgPSBNZW1vcnlQcmVzc3VyZUhhbmRsZXI6OmN1 cnJlbnRNZW1vcnlVc2FnZVBvbGljeSgpOworICAgICAgICBjYWNoZWRNZW1vcnlQb2xpY3lUaW1l ID0gbm93OworICAgIH0KKworICAgIG1fY29tcG9zaXRpbmdQb2xpY3kgPSBjYWNoZWRNZW1vcnlQ b2xpY3kgPT0gV1RGOjpNZW1vcnlVc2FnZVBvbGljeTo6VW5yZXN0cmljdGVkID8gQ29tcG9zaXRp bmdQb2xpY3k6Ok5vcm1hbCA6IENvbXBvc2l0aW5nUG9saWN5OjpDb25zZXJ2YXRpdmU7CiAgICAg cmV0dXJuIG1fY29tcG9zaXRpbmdQb2xpY3kgIT0gY3VycmVudFBvbGljeTsKIH0KIAo=