Performance:Home Page

This page is meant to provide a starting point for developers who want to improve performance in Gecko and Firefox.

This page was last updated March 18, 2010.

Dashboard

• Per-branch graphs over time
  • Firefox 3.5
  • Firefox 3.6
  • Trunk
• Cross-branch comparison snapshot

Projects

• Startup Performance
• Front-end Performance (i.e., responsiveness)
• Add-on Performance

Tools

• Codesighs - a tool which analyzes code and data size.
• Performance:Tools. For measuring performance. Tip o' the propeller-cap to Daniel Roberts (zuperdee@penguinpowered.com) for the pointers.
• Performance:Probes. Project to integrate a system of performance probes into Gecko.

Tests

The Performance:Tinderbox_Tests are commonly used as a benchmark of Mozilla performance.

Documentation

• Code Footprint Performance:Footprint_Reduction_Techniques explains common bad patterns and how to correct them.
• Performance:Profiling_JuJu. Things you should know about doing profiling. Tips and tricks for some of the tools, and lots of other Good Things To Know.
• Profiling and leak analysis on the Mac.

Tool Presentations

• Performance:Footprint_Tools. Presentation on footprint tools
• Performance:Startup slides Presentation on general performance tools
• Performance:Leak_Tools. Presentation on memory leaks detection tools
• Mac Performance Tools Presentation on performance tools that work on Mac

Task Lists

Somewhat ordered in descending probability of being useful, based on an inexperienced observer:

• bug 447581 - tracking bug: startup performance
• bug 71668 - tracking bug: page loading performance issues
• bugs with topperf keyword
• bugs with perf keyword
• bug 91351 - tracking bug: UI/App responsiveness issues
• bug 384115 - tracking bug: Mozilla 2 "exploiting hardware parallelism"
• bug 72677 - tracking bug: putting timer hooks in app code for performance measurment
• bug 191033 - tracking bug: embedding code size reduction
• tracking bugs: Cairo performance issues
  • Windows: bug 334719
  • Unix: bug 334720
  • Mac: bug 334721
• bug 384323 - tracking bug: UI responsiveness - core/platform
• bug 384325 - tracking bug: UI responsiveness - applications
• bug 98275 - tracking bug: xpcom perf
• bug 307312 - tracking bug: startup with WAY_TOO_MUCH_GC
• bug 500489 - tracking bug: Slow performance on this DOM manipulation testcase
• bug 499198 - tracking bug: Peacekeeper benchmark
• bug 377787 - tracking bug: Cycle collector performance
• bug 98284 - tracking bug: xpcom footprint
• bug 49141 - tracking bug: New window performance
• bug 103712 - tracking bug for Mozilla 1.0 performance

Old Stuff

First you could help us consolidate these old tasklists.

• An outdated Performance:Projects list.
• Tasks. A task list of hot-button items! See the Porkjockeys newsgroup for the latest on performance and runtime footprint. Pigs. Flight. And all that.
• Laundry List. Another grouped listing of all open performance bugs. (Wildly out of date. Needs to be updated!)
• Slightly more recent but still out of date list: PerformanceList

Tasks

The list below focuses on the performance projects that we think will have the biggest impact. Each task includes a brief description and links to additional resources. The tasks are grouped by performance area but are otherwise in no particular order.

Startup Performance

Reduce startup workload

We should be able to greatly reduce the amount of work needed to bring up the Firefox browser window. Today, it loads a bunch of extra stuff that really could be delayed until it is needed. Phoenix 0.1 started up very quickly because it was a simple XUL browser! ;-)

Possible areas for improvement:

• Put components which aren't typically needed at startup into external dlls, to reduce the amount of code that has to be paged in
• Fastload XBL bindings
• ...

PUT all in one File

NO extra DLLS and no DLLS because current harddrives load 40 to 80 MB per second but only if it is one big file and second manage internaly what classes are instantiated and loaded into workingmemory only whats needet to see the browsers userinterface. pleas make a version without script for users the dont care about (99%) --nonkapitalist 14:46, 16 August 2006 (PDT) --- Typically one would put the mandatory for first page functions in one DLL then divide up other functions into DLL's that can load quickly on demand. Needing to access multiple DLL's to display the first screen, I agree would generate unnecessary drag.

Move FF to XULRunner (investigate at least)

A trimmed down XUL browser application based on xulrunner loads much more quickly than FF 1.5, and uses significantly less memory (on the order of 14 vs 26 megabytes of memory). We need to understand why, and more importantly figure out how to trim the size of FF's memory footprint accordingly. This may mean moving FF to xulrunner, or it may mean applying something else that we learn from the investigation. At any rate, we know it should be possible to show the user's homepage using significantly less memory than we do now, which should help startup performance.

Preload FF at system start

Bring back Mozilla quick launch, but do it better ;-) Possibly build an extension that could be used with FF 1.5 to enable this functionality.

Javascript Component Loading

Even with fastloaded (precompiled) javascript components, we spend a fair amount of startup time deserializing large scripts. Things to do:

• Refactor large components (nsExtensionManager.js, nsUpdateService.js) to load the minimal amount of code at startup
• Eliminate locking in the JS engine when deserializing scripts, see bug 312238
• Use a large pool of memory for allocating strings during deserialization, see bug 279839#c11

Function Ordering

All of the major platforms support having the linker reorder functions in the binary. We can exploit this to move the functions that are needed at startup into a contiguous section at the start of the binary, which should reduce the time spent reading from disk.

• Profile the startup process to generate an optimal function ordering
• Integrate the ordering into the build system

---

• Might also want to see if anything in the startup sequence can be parallelized...

Improve Fastload system

The Fastload system is used to load precompiled javascript, preparsed XUL, and soon preparsed CSS at startup. The system is currently designed to be architecture-independent, which makes it impossible to write large chunks of memory at once (differences in endianness, data sizes, and struct packing). We could drastically reduce the time spent reading the fastload file by changing to an architecture-dependent, memory-mapped file that could be accessed directly.

Page load and DHTML Performance

Optimize XPConnect Locking

XPConnect does a lot of locking for threadsafety despite the fact that 99% of callers are on the main thread. We should optimize this for the common case. This should directly impact DHTML performance. For more information, see https://bugzilla.mozilla.org/show_bug.cgi?id=307953 .

Optimize for the main application thread

99% of the work done by Gecko happens on the main thread, and yet many of the core pieces of Gecko are written to be fully threadsafe. This includes core XPCOM components, XPConnect, and portions of Necko. There is also an overuse of threadsafe reference counting (using atomic increment/decrement). We can shift more of the performance burden on background threads by using proxy objects when accessing "main thread only" objects on background threads. We can do this today with some hardcoding of NS_GetProxyForObject, but it would be better if object proxification were automated. For example, getting a component from the component manager on a background thread should result in a proxy to the real object if the object only wishes to be invoked on the main thread. This will allow us to more easily streamline main thread object access while preserving support for those infrequent background thread use cases.

Request Prioritization

Necko has support for prioritizing requests to the same host relative to each other. We can make use of this to give on-screen content higher priority than off-screen content. Images, plugins, iframes, etc. could benefit from this.

Other areas for investigation

These ideas are less well-defined than the ones above, but may be good starting points:

• Reuse the scrollbars and associated objects between pages. Currently, these are recreated on each load.
• Experiment with the timing of initial layout and paint. We may be able to help at least perceived performance by trying to get a full screen of content displayed as fast as possible (but without "jumping").
• Optimize the way that HTML reflow works. See Gecko:Reflow_Refactoring.
• Investigate better strategies for allocating memory for DOM and layout objects.
• Other items listed at Gecko:Layout_Improvements.
• XPConnect currently makes JS -> C++ calls via xptcall, which is really an interpreter. Could we identify common method signatures and created compiled translation stubs for important platforms?

Mobile Related Issues

Mobile Performance Info

Other Related Projects

• (ARCHIVED) Mail-News Performance Statistics. A history of our mail-news performance stats.
• [http://www.mozilla.org/xpapps/performance XPApps Performance. The browser performance effort.
• Performance:References. A dumping ground for performance-related links and documents.
• Community. netscape.public.mozilla.performance is the Mozilla performance newsgroup, and mozilla-performance@mozilla.org is an e-mail mirror.