Meeting 2024-04-02

• Team calendar

  • Next on Necko triage: Valentin

• Round Table

  • Necko Performance Opportunities select priorities for H1

    • Performance priorities selected for H1-2024

      • HTTP3 Upload Speed

      • Enable DNS pre-fetch for https

      • Bug: 1709175 Switch from select() to poll for MacOS to reduce CPU usage

      • HTTP/2, HTTP/3 priorities, urgencies, incremental flag(meta bug)

      • Optimize socket pool constants - Android

  • [acreskey] FYI:

    • Fenix performance gap to Chrome through throttling VPN reproduced

    • Live pageload comparisons

    • https://share.firefox.dev/3TVlXT3

      • Problems with tls, general request times

• Active Projects roadmap 

  • Active Projects: Fx125 rel Apr 16, Fx126 soft freeze Apr 11 

    • OMT: Move OnStartRequest and OnStopRequest to the content process (paused, target: TBD)

      • OMT OnStart/OnStop Request (content) threading/api model

      • new Project Plan

      • Split HTML5Parser OnStop - OMT Shipped in Fx122 

      • Split CSSLoader OnStop - OMT to track in Fx126

        • Experiment started 

      • OnStart consumer

        • – enter updates –

    • Optimize, Measure and add tests for Speculative Connections (in progress, target: Fx126) 

      • [acreskey] Update:

        • Beta experiment live, ended Mar 19, 2024 - no measurable improvements via telemetry

        • Local browsertime tests show significant improvements to sub-resource connections times (up to 50%) on some sites

        • (quick demo)

        • We don’t hit socket pool exhaustion on Desktop:

          • https://bugzilla.mozilla.org/show_bug.cgi?id=1819556
            We removed the telemetry for desktop socket exhaustion

            • NETWORK_SESSION_AT_900FD probes.       Telemetry shows all users are able to poll up to 1000 sockets - see glam

        • Landing patch to set Desktop to 20 for Fx126

    • Native HTTPS RR Implementation (in progress, target Fx126)

      • [valentin] Enabled on Nightly and Early beta 125

      • Investigating issues:

        • 1885529 - Investigate increase in DNS queueing time with native HTTPS records (Windows)

          • Telemetry patch ready.

          • I can repro locally on win11 (investigating)

        • 1885556 - 4.62 - 4.32% fandom ContentfulSpeedIndex  fandom loadtime + 1 more (Linux) regression

          • No progress yet

    • Interop 2024 URL (in progress, target H2 2024)

      • [Ed] In progress: unknown scheme default URI bypass list 

      • [Dylan]

        • Landed https://bugzilla.mozilla.org/show_bug.cgi?id=1879317 

        • Submitted patch in rust-url https://bugzilla.mozilla.org/show_bug.cgi?id=1874117 

    • 3rd Party Cookie Deprecation (in progress, target Apr 9)

      • [Ed] reviewed telemetry for guidance on order of potential purging when reaching cookie partition byte count limit (FIFO vs FILO)

        • Currently it’s not even clear on if this should be enforced at all

        • No existing telemetry to indicate how often a limit would be reached and potentially interfere with web compat

        • We can probably proceed with multiple purging implementations behind feature flags and then decide based on some telemetry that ships ahead of it.

        • Any concerns/differing opinions?

    • zstd content-encoding: (in progress, target H1 2024)

      • Code r+’d to land

        • xz library exploit doesn’t affect our import since we don’t build zstd binaries

        • Likely will update the library anyways since a new release with performance improvements, etc was just made to coincide with Chrome enabling zstd in release.

      • Works well with facebook

      • In contact with decoder about fuzzing

      • Remaining work with sec team to TLS cert comp and dev tools

    • Fetch keepalive: (in progress, Target Fx128) 

      • Testing the initial implementation

    • Performance work: 

      • [jesup] Results are in from the OMT decompression landing: https://treeherder.mozilla.org/perfherder/alerts?id=41924 

        • Some red. Lots and lots of big green through

        • Filtering to speedindex looks even better. 100% green for FCP and FirstVisualChange. LCP is good, though not all green.

        • Pushed tries for a patch that leaves things on MainThread if the compressed size is known and is small (512 and 1024 bytes). This could avoid overhead in cases where decompression is almost certainly very short. 

          • 512 shows improvement of 1-7% on 4 sites (of 30)

          • 1024 shows improvement of 1-3.5% on 5 sites, regression of 1-3% on 2 sites.

          • Planning to land with 512; includes a pref for size