I would like to share some updates about some of the ongoing performance related work.
We have started looking at the native stack traces that are submitted through telemetry from the Background Hang Reports that take more than 8 seconds. (We were hoping to have been able to reduce this threshold to 256ms for a while now, but the road has been bumpy — but this should land really soon now!) Michael Layzell put together a telemetry analysis job that creates a symbolicated version of this data here: https://people-mozilla.org/~mlayzell/bhr/. For example, this is the latest generated report. The grouping of this data is unfortunate, since the data is collected based on the profiler pseudo-stack labels, which is captured after 128ms, and then native stack (if the hang continues for 8 seconds) gets captured after that, so the pseudo-stack and the native stack may or may not correspond, and this grouping also doesn’t help going through the list of native stacks and triage them more effectively. Work is under way to create a nice dashboard out of this data, but in the mean time this is an area where we could really use all of the help that we can get. If you have some time, it would be really nice if you can take a look at this data and see if you can make sense of some of these call stacks and find some useful bug reports out of them. If you do end up filing bugs, these are super important bugs to work on, so please make sure you add “[qf]” to the status whiteboard so that we can track the bug.
Another item worthy of highlight is Mike Conley’s Oh No! Reflow! add-on. Don’t let the simple web page behind this link deceive you, this add-on is really awesome! It generates a beep every time that a long running reflow happens in the browser UI (which, of course, you get to turn off when you don’t need to hunt for bugs!), and it logs the sync reflows that happened alongside the JS call stack to the code that triggered them, and it also gives you a single link that allows you to quickly file a bug with all of the right info in it, pre-filled! In fact you can see the list of already filed bugs through this add-on!
Another issue that I want to bring up is the [qf:p1] bugs. As you have noticed, there are a lot of them. 🙂 It is possible that some of these bugs aren’t important to work on, for example because they only affect edge case conditions that affects a super small subset of users and that wasn’t obvious when the bug was triaged. In some other cases it may turn out that fixing the bug requires massive amounts of work that is unreasonable to do in the amount of time we have, or that the right people for it are doing more important work and can’t be interrupted, and so on. Whatever the issue is, whether the bug was mis-triaged, or can’t be fixed, please make sure to raise it on the bug! In general the earlier these issues are uncovered the better it is, because everyone can focus their time on more important work. I wanted to make sure that this wasn’t lost in all of the rush around our communication for Quantum Flow, my apologies if this hasn’t been clear before.
On to the acknowledgement section, I hope I’m not forgetting to mention anyone’s name here!
- Bas Schouten made it so that we don’t clear the compositor background immediately before drawing into it. This made some painting and scrolling related benchmarks faster, and fixed a flickering issue in the mean time!
- Mason Chang made the Youtube settings widget less janky on Windows with D2D when the video is fullscreen.
- David Baron made the flushes caused by the code that watches your mouse to know which side of the window to put the status bar when you hover a link less severe.
- Neil Deakin removed a synchronous layout flush that used to happen when closing a window which would slow down the window going away.
- Dão Gottwald removed some obsolete tab animation telemetry which could slow down tab animations (yes!). Dão also removed a synchronous layout flush which could slow down detaching a tab into a new window and he also lazified some code to avoid some more layout flushes which were related to the status panel.
- Markus Stange made the Gecko Profiler not kill your battery after you stop profiling, and also made it so that once you start profiling a thread, you can actually stop profiling it.
- Michael Layzell excluded some common file JS file extensions from the expensive computation that we run on Windows to determine the mimetype of a file from its extension.
- Kris Maglione made it possible to load add-on SDK modules lazily. He also made loading ExtensionContent.jsm lazy. These two changes together (but probably mostly the former) showed great improvements on Talos. This is even true for users without any add-ons using the add-on SDK since these modules are also used in our internal code.
- Robert Helmer created a Go Faster add-on to quickly deploy the fix of bug 1353216 to Firefox 52 users even before they upgrade to Firefox 53 which includes the fix!
- Edouard Oger made Firefox Sync UI code cache the DateTimeFormat objects where possible.
- Alastor Wu made some Fennec media controls code that could be expensive not run on Firefox desktop.
- Makoto Kato added a cache for document encoder objects used by TextEditor, in order to improve the performance of setting the value of HTMLInputElement.value.
- Florian Quèze removed a synchronous layout flush that used to happen when displaying each item in the awesome bar. He improved the awesomebar even more by caching the one-off search buttons instead of regenerating them every time we open the popup. Furthermore, he ensured that we don’t pay the cost of initializing UITour.jsm for pages that do not display a UI tour.
- Wei-Cheng Pan added telemetry probes for navigation timings and time to first byte measures.
- J. Ryan Stinnett made the Developer Toolbar not create a network listener for every single tab that you have open. Besides perf wins, this means fewer OOM crashes too!