The Cloud Native Compiler: JIT-as-a-Service
summary
Cloud-native approach extracts the JIT into a centralized service so JVMs send bytecode plus profiling and receive optimized native methods back, cutting warm-up and local CPU use. Replacement compiler Falcon (LLVM-based) emits AVX2 vectorized loops and boosts tested workloads — Cassandra ~22% faster, JMH stream ~60%, Kafka ~40%, typical overall ~20% faster — while a shared fingerprint cache reuses speculative-optimized variants across instances. The system ships as a Kubernetes pod (Prime JVM integrates CNC), currently needs architecture-specific CNCs for x86 vs ARM, and it works in production as described.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 3:12JVM warm-up challenges
- 8:11Microservices and JVM limitations
- 11:19Cloud native JVM concept
- 13:04JIT compilation importance
- 19:29Speculative optimizations and de-optimizations
- 22:00Vector operations and Falcon JIT
- 25:51Trade-offs in better JIT
- 28:15Cloud native compiler benefits
- 31:09Centralized JIT efficiency
- 35:00AOT vs JIT performance
- 36:47Future optimizations
- 38:56Cloud native compiler summary
- 40:00Q&A
related talks
52mCloud Native JVM. Cloud Compiler
06mLightning Talk: A Quick Intro to Falcon
40mNext Frontier in Open Source Compilers: Just-In-Time Compilation as a Service
38mThe Next Frontier in Open Source Java Compilers: Just-In-Time Compilation as a Service
55mJIT and AOT in the JVM
33mNew Opportunities for Developers with GraalVM
43mPolyglot Native: Java, Scala, Kotlin, and JVM languages
43mThe Virtual Machine Under the Hood