frequent CMS collections/ CPU spike/ Hotspot JRE1.4.2_17/

[vasu ts]

Hi all,

We have an application which is deployed on IBM websphere 5.1/Solaris 5.9/Sun hotspot JRE1.4.2_17/. We have 4 JVM's which are running on the same machine. 
These JVM's recieve xml messages from MQ queue which are processed ( business logic stores the data from xml into database) and xml replies are sent back to the MQ queue.

Hardware
 8 dual core - sparc IV
 4 single core - sparc III

JRE options set on JVM's

-server -XX:+PrintGCDetails -XX:+PrintGCTimeStamps -Xmx768m -Xms768m -XX:MaxNewSize=500m -XX:NewSize=500m -XX:+UseParNewGC -XX:+CMSParallelRemarkEnabled -XX:+UseConcMarkSweepGC -XX:MaxTenuringThreshold=0 -XX:SurvivorRatio=128   

During our stress test we are seeing that CMS collector is trying to start old gen collection very frequently and the cpu usage spikes upto 99% -100% usage. 
Our stress test included adding a user per second until we reach 2500 user limit and then we maintain steady user rate of 2500.

Attached are the gc logs from one of the JVM, the PrintGCStats details and GCTimeline graph from GCHisto tool

Is there anything I should set so that CMS collector doesn't start so frequently. Also, don't know if increasing the total heap size (to 1GB) will improve this situation.

Please provide your comments.

thanks
vasu..


_________________________________________________________________
Windows Live™: Life without walls.
http://windowslive.com/explore?ocid=TXT_TAGLM_WL_allup_1b_explore_042009
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.openjdk.java.net/pipermail/hotspot-gc-use/attachments/20090420/5ba4c791/attachment.html 
-------------- next part --------------
A non-text attachment was scrubbed...
Name: nativelogs.zip
Type: application/octet-stream
Size: 222760 bytes
Desc: not available
Url : http://mail.openjdk.java.net/pipermail/hotspot-gc-use/attachments/20090420/5ba4c791/attachment.obj