Java performance
Download as PPTX, PDFā¢1 likeā¢1,974 views
This document discusses how to debug Java performance issues by considering key metrics like throughput, latency, and concurrent users. It outlines potential bottlenecks related to memory, CPU, disk I/O and network I/O. Specific techniques are provided to analyze memory issues using heap dumps and JSTAT, CPU issues using thread dumps, and disk I/O issues using I/O statistics. The document also briefly covers the different garbage collector options in Java.
Java performance
- 2. Few Questions to Consider ā¢ What is expected throughput ā¢ What is expected latency ā¢ Maximum No. of concurrent users/tasks ā¢ Expected Throughput/latency at max users ā¢ Worst case latency ā¢ Tolerable GC latency
- 3. Bottlenecks ā¢ Memory ā¢ CPU ā¢ Disk I/O ā¢ Network I/O
- 4. Memory (RAM) ā¢ -Xmx too low ā¢ Process leaking memory (et tu java!) ā¢ Insufficient DirectBuffers ā¢ Insufficient PermGen space ā¢ Page thrashing
- 5. Memory (Contd.) ā¢ Run āXX:+HeapDumpOnOutOfMemoryError. Check .hprof files in working directory or ā XX:HeapDumpPath ā¢ jmap āhisto<pid>
- 6. Memory (contd.) ā¢ High si/so coinciding with low āfreeā memory implies page thrashing
- 7. Other OutOfMemory Errors ā¢ Incorrect PermGen size āXXMaxPermSize ā java.lang.OutOfMemoryError: PermGen space ā jstat -gc<pid> ā Can monitory āPUā vs āPGCMXā from jstat - gccapacity ā¢ Incorrect Direct Buffer limit - XX:MaxDirectMemorySize ā java.lang.OutOfMemoryError: Direct buffer memory
- 8. CPU ā¢ Check Garbage collector stats ā Jstat āgc<pid> ā¢ Investigate using ā JMX, TOP ā Using TOP get the thread ID ā Thread Dump using ākill -3ā or ājstackā ā Search Hex thread id in thread dump
- 9. Disk I/O ā¢ Iostat %iowait<interval><samples> ā¢ High %iowait indicates I/O bottleneck ā¢ If java process is causing high I/O, check the thread dump
- 11. Garbage Collector Choice ā¢ Serial GC (Never use this!) ā -XX:+UseSerialGC ā¢ Parallel GC (High Throughput) ā Use in non- interactive jobs which can tolerate pause of 1- 2 sec ā -XX:+UseParallelGC āXX:+UseParallelOldGC ā¢ CMS (Low Latency) ā Good for webapps and interactive processes ā -XX:+UseConcMarkSweepGC