GC Optimization Problems

Question

We optimized our GC Parameters for JDK 1.4.2_06 to reduce GC time:
"-server -ms1024m -mx1024m -verbose:gc -Xloggc:/var/applications/BOL81/logs/inhouse1/gc.log -Xnoclassgc -XX:+UseConcMarkSweepGC -XX:+UseParNewGC -XX:+PrintGCDetails -XX:+PrintGCTimeStamps -XX:MaxPermSize=512M -XX:NewSize=384M -XX:MaxNewSize=384M -XX:SurvivorRatio=6 -XX:+UseLWPSynchronization"
After several hours under heavy load the GC occurs very often. Up to 4 times
per minute and last about 8 seconds.
Here is an extract of the GC log:
22800.298: [GC 22800.299: [ParNew: 392576K->392576K(392896K), 0.0000529 secs]22800.299: [CMS: 272361K->272804K(655360K), 7.3306386 secs] 664937K->272804K(1048256K), 7.3311409 secs]
22818.446: [GC 22818.446: [ParNew: 392576K->392576K(392896K), 0.0000532 secs]22818.446: [CMS: 272804K->278247K(655360K), 7.8801372 secs] 665380K->278247K(1048256K), 7.8806054 secs]
22842.256: [GC 22842.256: [ParNew: 392576K->392576K(392896K), 0.0000515 secs]22842.257: [CMS: 278247K->273676K(655360K), 7.8000559 secs] 670823K->273676K(1048256K), 7.8005780 secs]
Is this behaviour correlated with -XX:+UseConcMarkSweepGC or is
the NewSize just to big.
Sincerely
jpurgahn

Answer 1

It depends on what your program is doing. But it looks like your young generation is too
large.
The ParNew GC takes almost no time 0.000053 seconds but does not manage to clear
any memory. CMS (concurrentMarkAndSweep) seems to be freeing very little memory
and takes 7.8 seconds to do it (it frees about 0.5M in 7.8 seconds).
You might also want to try the -Xincgc to enable incremental collection of the old
generation. Note that this will not reduce the amount of time spent performing gc on
the old generation it just allows part of the old generation to be cleaned up on each
minor GC (can reduce the pause) (it tends to be slower overall)
The following link is useful
http://java.sun.com/docs/hotspot/gc1.4.2/example.html
You might also wish to add the -XX:+PrintTenuringDistribution option to see the
distribution of objects
or use the
-XX:+UseParallelGC -XX:+UseAdaptiveSizePolicy instead of -XX:+UseParNewGC
using +XX:UseConcMarkSweepGC can cause fragmentation in the tenured
generation causing larger numbers of major collections to occur then with the
XX:UseParallelGC.
lots of options.
matfud