Ludicrous Speed: Designing For Performance on the JVM
summary
Claims JVM now outperforms C++ in many low‑latency domains and a real system processes 2 trillion metrics per day. Concrete tools and results: use hppc to avoid boxed primitives (≈27 MB vs 62 MB for ~1M entries), generate Fast Tuple code to store contiguous heterogenous tuples (benchmarks show nanosecond field access and ~4× memory reduction versus lists of lists), measure with a Java agent/ObjectExplorer and JMH, and avoid JNI-heavy direct allocation because unsafe.allocateMemory is slow.
AI-written summary · about
chapters (11)
- 0:00Start
- 5:44In Relative Terms (Ins - 1mi)
- 8:053 Regions of Memory
- 9:07Primitives vs Objects
- 9:29Object Overhead
- 10:25Java's False Trade
- 20:57Memory Consumption Demo
- 24:24Microbenchmark
- 25:48JMH Example
- 31:32Observing The JIT
- 31:53Example Assembly Dump
