Performance tuning Twitter services with Graal and Machine Learning
summary
Twitter runs many services on Graal and uses Autotune (a Bayesian-optimization service) to pick three inlining-related parameters; swapping Graal for C2 cuts user CPU ~13% on one tweet service and Autotune finds configs that reduce CPU by about 8.4% in a 24h experiment and up to 18% in some runs, lower GC cycles by ~7%, and improve p99 latencies ~28%. The system is built as an Autotune backend that launches controlled 30-minute evaluations and enforces throttling constraints.
AI-written summary · about
chapters (14)
- 0:00Start
- 8:08PERFORMANCE TUNING
- 9:59OPTIMIZATION OF A BLACK BOX FUNCTION
- 13:37WE LIKE BAYESIAN OPTIMIZATION...
- 14:39WHAT IS AUTOTUNE?
- 16:14WHAT IS GRAAL?
- 19:03WHICH PARAMETERS?
- 25:09TEST SETUP
- 27:03TWEET SERVICE: OBJECTIVE
- 27:32TWEET SERVICE: CONSTRAINTS
- 38:22SOCIAL GRAPH: OBJECTIVE
- 38:40SOCIAL GRAPH: CONSTRAINTS
- 45:34WHAT HANGES?
- 48:33SUMMARY
related talks
34mPerformance tuning Twitter services with Graal and Machine Learning
49mPerformance tuning Twitter services with Graal and Machine Learning
46mPerformance-Tuning Twitter Services with Graal and Machine Learning
56mPerformance tuning Twitter services with Graal and Machine Learning
34mPerformance Tuning Twitter Services With Machine Learning, Part 2 JIT and GC
53mTwitter's quest for a wholly Graal runtime
50mTwitter's quest for a wholly Graal runtime
41mTwitter’s Quest for a Wholly Graal Runtime