Contention point in the AOP layer

[michael.yuan]

I cannot seem to be able to drive the CPU load to >85% in my Seam application stress testing. If I increase the client threads, the response time would simply increase and the CPU stays at around 85%. So, I did a thread dump, and found most HTTP threads are waiting at line 161 of class org.jboss.aop.metadata.MethodMetaData. This line is a synchronized access to a HashMap, and it is appreantly a contention point where multiple CPUs have to wait on each other.

public class MethodMetaData implements MetaDataResolver {

 public synchronized Object getMethodMetaData(Method method, Object tag, Object attr) {

 // This is line 161:
 SimpleMetaData methodData = (SimpleMetaData)methodMetaData.get(method.toString());

 }
}

Clebert suggested to use a concurrent HashMap instead of the method level synchronization here. What do you guys think?

[clebert.suconic]

Or if you need to synchronize two HashMaps, maybe manually add partitions to MetaData. This is definitely a contention point.

[clebert.suconic]

I will move this forum to AOP development.


This was during EJB3 tests, but this feature is specific from AOP.

[kabirkhan]

http://jira.jboss.com/jira/browse/JBAOP-269

I have replaced this with a ConcurrentReaderHashMap in head. Please let me know if this helps

[kabirkhan]

As requested, I have upgraded AOP in jboss-4.0.x to a 1.5.1.snapshot