Java Platform performance

Java Platform performance

Tim Cramer, David Dagastine, Ben Kim
Sun Microsystems

Focus: Understand how to measure performance.
Should performance be added as a step to the development process?

Why taking the average when using benchmarks doesn't work
2 runs score 24 and 26 vs. score of 1 and 49.
Both average 25

Beware Microbenchmarks!

Student's T-Test
Determine confidence between 2 data sets
Want 95% or higher confidence level, p-value 0.05

Modern Java Virtual Machine(JVM) Software Microbenchmark Negating Characteristics
Warmup period for Dynamic compilation
on-stack replacement
separate compile thread
disk cache
  On by default on on-solaris x86 OS-based systems
  Off by default on Solaris x86 OS-based systems

Statistical Gathering
  Monomorphic call transformation
  branch targets
  execution frequency

Background Compilation
  -Xbatch

Compiler Optimizations
  Dead code removal
  Inlining

Tips and Tricks
JVM Software Performance Tuning, pre 1.5
  Java hotspot VM is highly tunable

Select Server compiler with -server flag
Correctly size the overall heap
set min/max of heap to same value to minimize resizing
  +AggressiveHeap flag
    size of initial heap based on avail memory MIN(1/2 RAM, RAM-160M)
    flag can change at anytime!

Monitor JVM software using JVMstat tools

New garbage collectors
  Default
  Incremental
  Throughput

New Features of JDK 1.5
Smart Tuning
  Provide good out of the box performance without hand tuning
  examine host machine and configure hotspot appropriately
  dynamically adjust java hotspot VM software at runtime
  adaptive heap sizing policy during life of application.

client performance
  Class data sharing
    faster classload = faster startup time
    shared class archive, reduce footprint
    JFC/Swing and Java 2D API improvements

Overall
  know how to write a microbenchmark before doing so
  upgrade to jdk 1.5
  understand OS and hardware

http://java.sun.com/docs/performance