Perfetto: Occasional missing threadnames in traces |
||||||||
Issue descriptionExamples: https://bugs.chromium.org/p/chromium/issues/detail?id=900927 The symptom is that some metrics for some benchmarks will occasionally flake down to 0, as the metric is looking for trace events from specific named tracks in the trace and can't find them (the trace will only list TIDs, not names, for some of the threads). Example trace: https://console.developers.google.com/m/cloudstorage/b/chrome-telemetry-output/o/balls_svg_animations_2018-11-03_07-56-36_46552.html
,
Nov 9
,
Nov 16
,
Nov 16
The following revision refers to this bug: https://chromium.googlesource.com/chromium/src.git/+/c69179d52b614df11977d8481b4045fee62f4644 commit c69179d52b614df11977d8481b4045fee62f4644 Author: Eric Seckler <eseckler@chromium.org> Date: Fri Nov 16 22:02:43 2018 perfetto: Stop tracing on threads more reliably Moves resetting of TraceEventDataSource fields before flushing threads. This way, the thread's flush no longer races with the field resets. (If flushing occurs before the field resets, another thread may theoretically recreate its ThreadLocalEventSink before the resets, too.) Also flush the disabling thread in the same task, as a future task may be executed on a different worker thread. Bug: 903449 Change-Id: I13eb0d0a330ec66ffd826791f842c6fa7f8bd717 Reviewed-on: https://chromium-review.googlesource.com/c/1340325 Commit-Queue: oysteine <oysteine@chromium.org> Reviewed-by: oysteine <oysteine@chromium.org> Cr-Commit-Position: refs/heads/master@{#608972} [modify] https://crrev.com/c69179d52b614df11977d8481b4045fee62f4644/services/tracing/public/cpp/perfetto/trace_event_data_source.cc
,
Nov 21
Given crbug.com/907067 and crbug.com/907069 , I'm now pretty convinced that flush on worker threads is currently necessary - and that the above patch should fix the missing thread names issue. There's no way to verify this right now though, because the android webview bots that we saw the issue on most often are currently inactive / failing, see crbug.com/903204
,
Nov 27
,
Nov 29
It sounds like this is still happening for rendering.desktop benchmark when running locally. Is there some way to debug/diagnose the issue?
,
Nov 29
,
Nov 29
Everything is back to normal now so I think this bug is fixed. Thanks!
,
Nov 29
The missing thread names issue should be resolved (thought I'm still waiting for new benchmark results to verify), but events from some worker threads may be still be missing, see issue 907067 .
,
Dec 3
Looks like one of the very flaky graphs has returned and was stable over the weekend. |
||||||||
►
Sign in to add a comment |
||||||||
Comment 1 by eseckler@chromium.org
, Nov 9