Devoxx 2009: Performance for the Performance-shy

19/11/2009, Holly Cummins, IBM

costs of bad performance

• electricity
• employee productivity
• lost business (e.g. slow pages)
• hard cash (e.g. trading)

tuning technique

• find the bottleneck in the application
  • cpu → consistently high
  • I/O → cpu not consistently high
  • lock → cpu not consistently high
  • space ??? cpu can be low or high ...
• what is the JVM doing? → tools … must have:
  • visual indicators
  • explanation
  • suggested solution (command line)
• measure performance
  • risk to make it worse
  • get a baseline
  • system must be as similar as possible to production
  • how:
    • System.currentTimeMillis()
    • warm up your system
      • JIT
      • trigger caches

tools: IBM Perfomance Tools

• for IBM jvm's
• free
• IBM Support Assistant: centralized repository for IBM tools

space-bound applications: heap

• diagnose:
  • lower throughput
  • crash (out-of-memory exception)
  • trigger gc's → higher cpu
• tools:
  • verbosegc: very low overhead → use on production
  • live memory monitoring → IBM Monitoring & Diagnostic tools for Java
    • GC and Memory Visualizer
      • for Websphere → also works on Solaris & HP-UX (i.e. sun jvm syntax)
      • heap details, including heap histogram
      • get dominator tree
      • search inefficiently used data structures & redundant data
      • recommendations
        • increase heap size
        • gc takes between 10% - 20% cpu → tune gc policy
    • Health Center
      • agent in JVM
      • client (IBM Support Assistant)
      • visualize gc
      • recommendations
      • check gc performance
  • Memory Analyzer (not from IBM → eclipse.org): diagnose footprint issues:
    • heapdumps
      • automatically on out-of-memory
      • trigger programmatically
      • tool support is essentialµ
  • OS-tools
    • track native memory: perfmon, vmstart, ps (os-specific)
    • use gcmv to visualize os-specific dump
    • some native structures have java-wrappers in the heap (e.g. Threads and NIO buffers)

cpu-bound applications

• target your optimizations → measure
• diagnose:
  • method trace:
    • System.out.println()
    • log4j
    • ibm-jvm param -Xtrace:print=mt,methods=(HW*)
  • method profiler:
    • health center: low overhead sampler

I/O bound applications

• gc and memory analyser → long gc times might indicate paging
• method trace for network and disk I/O
• os-specific tools

synchronization issues (concurrency)

• sample thread dumps
• health center:
  • lock analysis
  • identify contended locks