Twitter's quest for a wholly Graal runtime

on YouTube ↗54m

" summary

Twitter replaces HotSpot C2 with the Java-written Graal compiler on core Scala services, runs it on a dedicated cluster and deploys it to production tweet and user services, and reports 11% CPU saving versus C2 with 2.5% fewer scavenge cycles (additional inlining tuning adds ~2% more). Bugs found during rollout are fixed upstream; Graal needs JVMCI/JDK9 support or a JDK8 backport. Tradeoffs: about 40 MB JVM heap overhead plus ~10–30 MB metaspace, and gains are larger for Scala workloads than for plain Java.

AI-written summary · about

" chapters (6)

1. 0:00Start
2. 0:58Goal?
3. 12:11Bugs
4. 22:18Contribution
5. 24:47Numbers
6. 34:19Is it escape analysis?

" related talks

• 53mTwitter's quest for a wholly Graal runtimeChris Thalinger · JavaDay Ukraine 2017
• 50mTwitter's quest for a wholly Graal runtimeChris Thalinger · Voxxed Days Ticino 2018
• 41mTwitter’s Quest for a Wholly Graal RuntimeChristian Thalinger · Oracle OpenWorld 2017
• 56mPerformance tuning Twitter services with Graal and Machine LearningChristian Thalinger · Voxxed Days Cluj-Napoca 2019
• 40mScala ♥ GraalFlavio Brasil · QCon São Paulo 2019
• 41mGraal: How to use the new JVM JIT compiler in real lifeChris Thalinger · Scale By The Bay 2018
• 46mPerformance-Tuning Twitter Services with Graal and Machine LearningChris Thalinger · Oracle Code One 2018
• 23mAdopting GraalVMPetr Zapletal · Scale By The Bay 2018