Twitter’s Quest for a Wholly Graal Runtime
summary
Twitter reports running the Graal JIT compiler in production for core Scala services (Tweet, user, social-graph), finds and fixes real crashes caused by agents and static-initializer edge cases, ports ~1,000 lines of handwritten intrinsics, and measures concrete wins: about 11% CPU reduction for the tweet service and roughly 2.5% fewer GC cycles, plus a small metaspace/runtime-state cost. Benchmarks include JDK 8/9/10 setups, inlining tweaks that add ~2% more, and plans for automated tuning and wider rollout.
AI-written summary · about
chapters (43)
- 0:00Intro
- 0:41Goal of this talk
- 0:54Chriss background
- 4:24What is Twitter
- 6:47Why run on Graal
- 11:27Compilation issues
- 15:24The fix
- 15:35The crash
- 16:31What is snippets
- 17:33Time to bite the bullet
- 17:43This never happened
- 17:57This was an important part
- 18:55Possible failure
- 19:31Twoline fix
- 20:32Compact Strings
- 21:47Integral
- 24:57Graphs
- 26:32Memory Leak
- 27:28Use CPU Time
- 28:10CPU Utilization
- 28:50Tweet Service
- 29:38Escape Analysis
- 30:51Money
- 31:07Data centers
- 31:21Cloud providers
- 31:49Cost per CPU core
- 32:38Cost of 10k servers
- 33:18What if we inline more
- 33:38Tweak this
- 33:54Second experiment
- 35:22JIT compiler
- 35:53Oracle JDK
- 38:05Java Conferences
- 39:33Summary
- 41:01Hands Up
- 41:13First Row
- 42:38Use Cases
- 43:27Inlining Parameter Tweak
- 43:40Optimizations
- 44:42IO
- 45:51Enterprise Crawl
- 48:40The Question
- 48:58The Answer
