Performance/2013-07-11

Performance Program Review - July 2013

Active Projects

Project	Status	ETA (weeks)
Incremental Cycle Collection (ICC)	rebased over other ongoing CC work investigating and fixing leakiness and crashes	Q3
Keep JS-accessible APIs from blocking the main thread	done: Async mozStorage connections Module loader for chrome workers Various tweaks to OS.File in progress: Workers Refactoring OS.File for module loader ETA: End of July for the next batch ETA: End of August for everything, unless something high-priority happens Storage Closing async mozthreadStorage connections off the main ETA: End of July APIs Async downloads API for JavaScript Async module for detecting crashes No ETA (mentored bug) Refactorings Improving asynchronousness + performance of SessionRestore startup blocked on reviews by ttaubert, who has some other regressions and other work to complete Adapting various modules to new-style Promise Making mozStorage clients async To begin Investigate DOM Promise ETA: End of August, unless something higher priority happens Port modules to async mozStorage connection	Workers: End Aug, mozthreadStorage: End July
Profiler Backend for Mobile	done: Multi-thread support for GUI (Anton Kovalyov) Multi-thread support for backend (BenWa + others?) Java profiling on Android (BenWa?) Native unwind for ARM/Android using EXIDX (in nightlies) at circa 200Hz / 100MB In progress: Improve breakpad unwind performance for EXIDX and CFI (jseward) <- Android and possibly Linux Have an in-principle plan, but don't know how hard it will be to implement Reduce mem use for EXIDX/CFI unwind. Closely related to previous point. Fix stability problems related to breakpad/faulty.lib interface (glandium, bug 887685) glandium holds the token on this; not sure how hard it will be to fix Other stuff we should consider Testing	Breakpad unwind speedups: late august.
Reducing time to first paint - Phase 1	The project hasn't reached the 10% improvement goal, partly because of regressions and partly because Vladan's time hasn't been devoted exclusively to this project. Retargeting for the end of July. done: set up an environment to collect consistent startup times, fixed Mozilla code that triggered stat-ing of plugins at startup, elminated loading of feclient.dll on startup https://blog.mozilla.org/vdjeric/2013/06/25/a-quick-firefox-startup-update/ note: cold time to firstPaint has improved almost 30% over the last year, Firefox 13 vs Firefox 21 (unrelated to my work) todo: reduce number of DLLs loaded at startup revisit decision to prefetch gkmedias.dll to dependentlibs.list (it gets read on startup) Improve SafeBrowsing behavior on startup (after firstpaint) new way to do startup crash detection that avoids main thread fsyncs	End of July
Replace Addon Manager SQLITE with JSON file	All core work is done and either reviewed or up for review Try builds exposed a small number of platform-specific test failures; fixes for those should be done this week Waiting for reviews on several key patches.	AddonRepository should land in the next few days, XPIProvider hopefully within a couple of weeks
Smooth Tab Animation	Newtab preload turned on by tim. Talos tab animation test: in progress (works, needs refinements - will leave some to v2). Scroll smoothness test: in progress (works, looking for better metric to correlete to smoothness).
Network Cache Rewrite	Testing on old iMac and Samsung with Android. Blog post with results from testing and optimizing on this hardware http://www.janbambas.cz/new-firefox-http-cache-backend-implementation/ Stuff not mentioned in the post: shutdown time regression due to closing all the files before data has been written gum needs to merge with m-c to get rid of a page reload crash on Android Notes: There is clearly a regression for cold cached load. Block files perform much better. We may consider to create bundle files and not just single files per entry. Needs definitely more discussion and testing. I'm using the event tracer to check on how stuff behaves - it's priceless to me! I can see that time to first paint is kinda volatile. It triggers randomly from 5 seconds to 30 seconds when doing a cold cached load (see the post). We may need similar resource scheduling as we have for network also in the cache code. Also, there could be a bug in thread implementation. Sometimes events are just hanging in the queue and not being triggered for e.g. 13 seconds, but that is a theory since I may not monitor all events happening on the given thread. More work on the cache I/O thread is definitely needed. Maybe just the better scheduling will revert that large regression.	Q4
Remove localStorage From Browser Pages	removed usage from all about: pages currently investigating possible regression of about:home (bug 890409) made patch to disallow using DOM Storage in about: page blocked on failures in Social and Jetpack, currently being worked on (bug 890089, bug 891516)	End of July
Complete Asynchronous History API	investigating an history regression caused by the new GetPlacesInfo API (bug 892454) Mano working on consumers conversion from GetPageTitle to GetPlacesInfo Writing a new Places transactions manager to take async into account for undo/redo of bookmarks restarting the work on RemovePlaces	Mid Sept
Reduce Impact Of Bookmarks Backups On Shutdown	https://etherpad.mozilla.org/places-backups-changes Raymond fixed a regression that caused backups restore to be very slow (bug 885205) bug 818593 blocked on a test failure due to file locking, undergoing investigation next is work to reduce number of backups through hash comparison	Late Sept
Reduce Storage Connections Main-Thread Operations	Yoric working on off-main-thread asyncClose, blocked on complex database close sequence causing leaks. Investigation ongoing Marco working on conversion of autocomplete to the async API, to avoid jank on first locationbar search Note: Consumers need to be converted after API is complete.	API: End July
Asynchronous Places Annotations	Mock of the API filed, Mano will implement prototype module on hold while working on GetPlacesInfo consumers conversion Note: Consumers need to be converted after API is complete.	Components: Late Sept
Improve Text Performance	Lots of fixes came in with bug 863248 - new shaping engine with a bunch of perf fixes More to come from bug 834609 (improve talos tests to be less noisy wrt text), bug 885689 (checks to ensure rtl text doesn't regress perf for ltr text) and others need to start with metrics so that we can prioritize bugs jdagget, jq, simon on this project	Q3
Downloads API rewrite	Work completed: Support for serializing to JSON file and file format definition (bug 835885, bug 851454) Methods to remove downloads when the related history entry is removed (bug 836485) Support for the becauseBlocked property on DownloadError (bug 852599) Next, only two bugs left to reach the "consumer interface complete" milestone: Add the ability to resume a download from where it stopped (bug 836437) Add methods to execute actions on completed downloads (bug 852482) Work to reach the "initial features" milestone (tracked in bug 881062) is proceeding in parallel	Mid Aug

Project Proposals

Nothing to report.