Java at Speed: Getting the Most Out of Modern Hardware
summary
Shows concrete gains from modern CPUs and JVM work: vectorized loops on AVX2 use masking to get ~6x speed and AVX‑512 runs a tuned summing loop at ~16–17 billion adds/sec on a 2.3GHz core; Falcon JIT generates CPU‑specific code, and speculative optimizations (class hierarchy, null checks) are validated and deoptimized at runtime. Covers microbenchmark pitfalls (JMH doesn’t prevent all dead‑code folding), a record/replay JIT warmup to avoid cold‑start profiling, and the C4 collector in Zing that largely removes GC pauses for smoother Cassandra performance (roughly 11% faster plus far fewer glitches).
AI-written summary · about
chapters (28)
- 0:00Intro
- 1:57Agenda
- 4:25What is speed
- 5:57Speed in Java
- 8:50Modern Servers
- 11:52Architecture
- 12:48Caches
- 16:49Machine Code
- 25:03Compiler Tricks
- 25:32Dead Code
- 25:53Propagation
- 27:26Cache Reads
- 29:34Inlining
- 30:59Loops
- 34:56Takeaways
- 36:12Speculative Optimization
- 39:01Arcane Analysis
- 40:40Chameleons
- 41:41Optimization
- 43:37Warmup doesnt cut it
- 45:43JVM replay
- 46:48What does this mean
- 47:24C4 Garbage Collector
- 47:39JVM Glitches
- 47:52Tuning
- 48:23GC Tuning
- 49:07Cassandra
- 49:59Conclusion
related talks
35mJava at Speed: Building a Better JVM
43mHow to make your electrons do more in Java?
44mJava at Speed: Building a Better JVM.
48mJust-In-Time Compilation for Java Performance: Recent and Ongoing Improvements
46mJava vs CPU
40mHow to make your electrons do more in Java
41mUp and Away: JDK Optimizations
48mJava 9 Performance