Re: BHR Project Status
> > > > Calixte made a tool to associate a newly reported crash to a recent > > changes. Would it be possible to look at the same kind of tools based on > > the reports of the BHR, with a given buildId ? > > > > Do you have a link to this tool, so I can understand a little better? > > This tool (see [1]) is mapping the files which appear in a crash backtrace (available on socorro) and the patches made in the last 3 days (according to the pushdate in hg.mozilla.org). So it helps to identify the patch causing a regression (see metabug [2]). [1] https://github.com/calixteman/clouseau/blob/master/clouseau/guiltypatches.py [2] https://bugzilla.mozilla.org/show_bug.cgi?id=1396527 > > > [1] https://arewesmoothyet.com/?category=all=2048_65536_ > ___ > dev-platform mailing list > dev-platform@lists.mozilla.org > https://lists.mozilla.org/listinfo/dev-platform > ___ dev-platform mailing list dev-platform@lists.mozilla.org https://lists.mozilla.org/listinfo/dev-platform
Re: BHR Project Status
On Wed, Sep 20, 2017 at 9:06 AM, Nicolas B. Pierron < nicolas.b.pier...@mozilla.com> wrote: > > What impact has a stack which is being reported at 1% / 0.5% / 0% ? > > I see that the histograms on top are changing each time I highlight a new > line. When I mouse-over the histograms buckets I notice the "ms/h" metric. > > What does it mean? Does that mean that in any browsing session, any user > will see the given signature for X ms per hour? > As a mean value, and for the date in question, yes. If a particular prefix list of stack frames is responsible for one 20 second hang for one in 10 users every 5 hours (pretending all users use Nightly for the same number of hours per day), you'll see a value of 20,000 / 10 / 5 == 400ms/h. > > Should the left column of the profiles report the time in ms/h instead > of unit-less % ? > That's a good suggestion. I'll work on getting units in there. > > > Is the histogram indexed by build-id / report-date? > Build ID. > > Is there a way to get the URL on which these hangs are reported? > If I'm understanding your question correctly I believe that's a privacy issue. Maybe if we end up going with RAPPOR we'll be able to see some aggregate stats on hanginess of particular domains though. Or we could prompt Nightly users to opt into sending more sensitive information, but that brings biases with it. > > > Can we have a larger history of hangs per signatures? (similar to > crash-stat [1]) > I'm working on keeping a dataset which will always go back to Sept. 1st. [1] I think we should be able to maintain this, however the performance of the UI will need to be addressed in some way, as even now it's a bit of a hog. > > Calixte made a tool to associate a newly reported crash to a recent > changes. Would it be possible to look at the same kind of tools based on > the reports of the BHR, with a given buildId ? > Do you have a link to this tool, so I can understand a little better? [1] https://arewesmoothyet.com/?category=all=2048_65536_ ___ dev-platform mailing list dev-platform@lists.mozilla.org https://lists.mozilla.org/listinfo/dev-platform
Re: BHR Project Status
On 09/20/2017 01:52 PM, Michael Layzell wrote: Doug Thayer has written a visualizer for the collected data called hangs.html (https://arewesmoothyet.com), based on the perf.html profiler viewer. This interface allows analysis of the change in frequency of specific hangs over time, lots of tools for filtering through hang information, as well as a profiler-like interface for drilling into specific hang stacks to determine what might be causing the problems. This looks like a really interesting tools, and I am glad we have this information in a much more readable format now! While I think I got a rough idea of the content which is presented, I want to make sure I read this content correctly. Thus, I have a few questions/comments, related to reading these data, and also related to investigating these reports, and to determine the importance (impact/priority) that we should give to each of these issues. > What impact has a stack which is being reported at 1% / 0.5% / 0% ? I see that the histograms on top are changing each time I highlight a new line. When I mouse-over the histograms buckets I notice the "ms/h" metric. What does it mean? Does that mean that in any browsing session, any user will see the given signature for X ms per hour? Should the left column of the profiles report the time in ms/h instead of unit-less % ? > Is the histogram indexed by build-id / report-date? > Is there a way to get the URL on which these hangs are reported? > Can we have a larger history of hangs per signatures? (similar to crash-stat [1]) Calixte made a tool to associate a newly reported crash to a recent changes. Would it be possible to look at the same kind of tools based on the reports of the BHR, with a given buildId ? [1] https://crash-stats.mozilla.com/signature/?product=Firefox=57.0a1=RtlEnterCriticalSection%20%7C%20mozilla%3A%3Anet%3A%3ACacheStorageService%3A%3ACacheQueueSize=%3E%3D2017-03-20T15%3A55%3A32.000Z=%3C2017-09-20T15%3A55%3A32.000Z#graphs -- Nicolas B. Pierron ___ dev-platform mailing list dev-platform@lists.mozilla.org https://lists.mozilla.org/listinfo/dev-platform
BHR Project Status
In the last few months we've been putting work into making the data which we collect from the Background Hang Reporter (BHR) more usable and actionable. We use BHR to measure the frequency and cause of browser hangs (when the main thread's event loop doesn't process events for 128ms or longer). The goal being to collect information which lets us improve Firefox's responsiveness by reducing the frequency of main thread hangs. On the data collection side, the BHR stack walking code has been rewritten to take advantage of Gecko Profiler internals. This reduced code duplication, and enables us to take advantage of Gecko Profiler features like JS stack interleaving. In addition, the ping submission logic has been rewritten to perform less work on main thread, and submit hang information outside of the main ping. This let us began collecting much more data, including interleaved chrome-js/native stack frame information for all hangs, and information about the browser's state, such as pending input events. Platform support has also been expanded from win32 to include linux64, win64 and macOS. Doug Thayer has written a visualizer for the collected data called hangs.html (https://arewesmoothyet.com), based on the perf.html profiler viewer. This interface allows analysis of the change in frequency of specific hangs over time, lots of tools for filtering through hang information, as well as a profiler-like interface for drilling into specific hang stacks to determine what might be causing the problems. Doug is actively working on adding new features to the UI to improve filtering and make it easier to get good results from the data, but we're already finding and fixing important bugs. Some bugs which have been fixed include bug 1393597 where we discovered that a synchronous GC on an edge case was having more of a performance impact than we expected, and bug 1381465 where we observed and prioritized the fixing of main thread I/O in the content process. ___ dev-platform mailing list dev-platform@lists.mozilla.org https://lists.mozilla.org/listinfo/dev-platform