SERVO: TODAY & TOMORROW

Jack Moffitt // Mozilla Research // September 26, 2016

Goal

Create a new browser engine making a generational leap in performance and robustness.

Performance Through Parallelism

Robustness From Rust

• WebAudio implementation in Firefox had 34 security critical bugs
• All were array out of bounds or use after free errors
• All would have been prevented by the Rust compiler

Servo Nightly - download.servo.org

Servo Nightly Reception

“It's definitely slower to actually load pages — I'm guessing that there are still a lot of network optimizations that more mature codebases have accrued that Servo hasn't yet — but holy cow are pages buttery smooth once they do load (and even while they're loading, which is unusual). Comparing Chrome and Servo in terms of UI jank felt pretty shocking, in Servo's favor. Kudos, it looks like Rust and WebRender have paid off.” - reissbaker

The Branches

• Parallel layout
• Parallel styling
• WebRender
• Constellation

PARALLEL STYLE & LAYOUT

Servo vs. Gecko

Servo vs. Gecko

Power Performance

• Running cores at a lower voltage reduces performance by 30% but power usage by 40%
• Servo can make up this performance loss through parallelism, achieving the same performance with 40% less power

Stylo

• Goal: Replace the CSS styling system in Gecko with the parallel version from Servo
• Initial experiments showed same performance gain when used in Gecko
• Telemetry shows ~95% of Firefox users will get at least 2x speedup, and ~50% will get at least 4x.
• First major uplift of Servo technology into Firefox

Stylo on Wikipedia

Stylo on HTML5 Spec

WEBRENDER

WebRender Motivation

WebRender in a Nutshell

• Hard to optimize immediate mode
• Retained mode graphics are better suited to GPUs
• Web pages are basically scene graphs, which can be sent to the GPU all at once
• Parallel CPU side preprocessing for some things (batch creation, font rasterization, border-radius, etc)

WebRender Progress

• Landed in Servo
• Optimized and fixed bugs
• Shipped in Servo nightly in June
• New design prototyped
• New design finished and landed

CONSTELLATION

SINGLE PIPELINE

PIPELINE WITH CROSS-DOMAIN IFRAMES

PIPELINE WITH SYNC IFRAMES

BENEFITS OF PIPELINES

• iframes don’t block their parent page or each other
• Layout and JavaScript don’t always block each other
• Many failures can be handled independently

MULTIPROCESS CONSTELLATION

The New Branches

• Constellation
• Parallel layout
• Parallel styling
• WebRender
• Magic DOM
• Parallelize MORE

Magic DOM

• Explore ways to achieve significant performance wins in JS / DOM integration
• Discussed for years, but a concrete proposal presented last year seemed fairly promising
• Fusing reflectors and DOM objects
• Self hosted JS

Parallelism Experiments

• JPEG decode (GPU)
• PNG decode
• Glyph rasterization (GPU)

External Research Collaborations

• Machine learning (Rohit Zambre at UC Irvine)
• Power management (Connor Imes at U. Chicago)
• Incremental computation (Kyle Headley at U. Colorado)
• WebBluetooth (Gabor Loki et al, University of Szeged)
• Software transactional memory (Alan Jeffrey)

QUESTIONS