TPEPerformance

Taipei Performance Team

This page is for tracking tasks of the performance team at Mozilla Taipei Office.

2014 Q2 Plans

• IPC with Async IO
• Nested Processes
• JS/GC/CC
  • Compacting GC
    • When will exact rooting be enabled safely?
  • GC heap analysis
    • How many and what pages/entries are touched?
  • the runtime XPCOMPtr information as complementary info. of the CC.
  • Instruction selector
  • Instruction scheduler
• session restore of WEB Apps
  • to get better memory usage and be flexible
• Necko
  • WEB cache
• Nuwa
  • Fix the IPC issues of preload-slow-things at the Nuwa process.
    • waiting for quiescent moment of the b2g process.
• Tarako
  • Priority manager

Working On

• Memory
  • Nuwa
    • NuwaTemplateProcess
    • Bugs
  • script-sources
    • In b2g and homescreen it consumes 4MB ~ 5MB totally.
    • In facebook and twitter it consumes ~1.3MB for each.
    • Need to investigate the benefits of disabling it under zRam.
    • Bug 944659 - Considerable amounts of script-sources
  • Misc
    • Bug 941792 - Make sure decommitting actually works on B2G
    • Bug 944224 - Considerable amounts of images as dataURLs
• IPC
  • Bug 915733 - Linux/or POSIX aio instead of IO thread
  • shmem for big data chunks
• Launch time of Apps
  • Preload JS code with Nuwa
  • Precompile JS code with the baseline compiler
  • Profiling
• Animations
  • use caches for display lists to reduce overhead of nsIFrame::BuildDisplayList()
    • https://groups.google.com/d/msg/mozilla.dev.tech.layout/BL3BNQu4sSA/CMHWMEHp9x0J
    • https://bugzilla.mozilla.org/show_bug.cgi?id=908626
• Task Tracer
  • https://bugzilla.mozilla.org/show_bug.cgi?id=908995
  • https://groups.google.com/d/msg/mozilla.dev.b2g/dr3RrYUvP-w/Q51za7NvIWwJ
  • Keep the tracks of tasks
  • TODOs
    • Bug 916409 - data retrieving tool and protocol
    • Bug 916410 - Integrate it with Cleopatra
    • Mark event sources
• XHR Batch Queue API
  • https://wiki.mozilla.org/WebAPI/XHRBatch
  • Submit a XHR to a queue on the device, the platform will handle the XHR until it is completed
  • Notify the App while all XHR on the queue is completed
  • The app can check its own queue
  • The app can be killed, or being stopped by the user, while XHR is still running.
  • The platform will wake up the app while the queue is completed an the app is not running.
• Stability Issues
  • Bug 926734 - [Flatfish][Profiling]: When starting up Gecko profiling tool, system got hang.
  • Bug 922548 - [Flatfish][Profiling]: Fails to analysis the parsed file for Gecko Profiling tool.
  • Bug 940851 - Gecko IPC hangs during boot on jb-gonk.
  • Bug 808080 - IPC parent aborts due to too large allocation for infallible array in PContent::Msg_AsyncMessage.
  • Bug 807738 - IPC parent aborts with malformed PHttpChannel::Msg_AsyncOpen.
  • Bug 781594 - ABORT: not reached: ipc/ipdl/PLayers.cpp, line 4069.

Ideas

• log size of gc heaps after every running JS runnable.
• Track JSObjects that fallback to dictionary mode
  • Time spent(impaction) on objects of dictionary mode
  • For games
• command queue for SVG frames
• Compare loading time for loading from scratch and from dumped image.
• Establish a benchmark to measure memory usage. Refer to AWSY.
• Bug 945166 - Merge JS bytecode, jit code, type tree and shape tree of iframes of embedded YouTube videos.
  • Ion/baseline codes, bytecode, script-source, type, shape, etc.
• Bug 945161 - Stop decode big image for the browser app for low end devices.
• Bug 945152 - Make memory of TypedArrays returned by XHR backing with the local file.
  • Use-case tables of keyboard apps.
  • Tim is in charged of the use-case.
• Bug 945174 - Evaluate zram performance and Bug 899493.
• Bug 945185 - Show only first N images in image uploader for low-end devices.
• Bug 943254 - [B2G][Browser] limit the memory consumption of the browser.
• Bug 854795 - decode and downsample images at the same time for memory savings
• App session resuming
  • like hibernate
  • load an App from storage and restore to the states of last killed.
• Media
  • Unified buffering
• Evaluate the code generator for ARM in IonMonkey.
• Parallelize layout and rendering by screen frames. -- VersionedDOM
• Use object layout info (size of the object) to figure out what the object is a XPCOMPtr in.
  • To be a complementary information for CC and GC.
• Tool to log and show restyling
  • What rules are referenced/matched?
  • Restyle roots and how many descendants?
  • Changes of DOM tree
• Automatically generate marionatte scripts
• build test servers

Current Status

• http://standu.ps/project/perf-tw

128MB

• Thrashing (GC <--> zRAM)
  • do experiments for proving
  • Solutions
    • stop GC
    • GC before switching to background
      • stop GC for memory pressure of background process.
  • Facts
    • memory pressure trigger GC, it brings pages from zRAM.
    • GC and memory minimize are called for switching to background.

Bugzilla query error

Array ( [type] => error [message] => http-bad-status [params] => Array ( [0] => 429 [1] => Unknown Error ) ) 1

Extending automated performance testing converage

• Proposed test cases to extend performance coverage.
  • Suspend/resume

    Time: receiving "power button" event -> first frame

  • Boot time

    Time: initrc -> lock screen ready. (visually loaded)

  • Scrolling response time

    Time: receiving touch event -> first composition that reflects the scrolling.

  • Long drag/Contineously scrolling.

    During drag, providing a list of composition time. (will be a series of events), provide median and standard deviation.

  • Rotate

    Time: receiving rotate event (i.e., the motion has been recognized as a rotation) -> first redraw that reflects the rotation.

  • Pinch

    Time: gesture being recognized -> first that reflects the pinch.

  • Button click

    Time: touch event -> first redraw of button's look change

  • Screen transition, after clicking on a button

    Time: action -> to the first frame of transition.

Trash

see https://wiki.mozilla.org/TPEPerformanceTrash