From your question I understand that you use a single cache and you don't restart it, but from the snapshot you attached, it looks like thread __ejb-thread-pool13 accessed at least 20 different caches. If it would be possible to keep using the same cache, without restarting it, I think that should solve your problem.

Note that ISPN-2372 is only relevant when the cache is restarted often: on restart, a new AbstractJBossMarshaller instance was created, but the thread-local PerThreadInstanceHandler instances were not discarded.

On the other hand, it looks like the fix for ISPN-2372 was indeed incomplete. We now clear the references from the thread-local marshallers to the cache itself, so it can be garbage-collected, but we don't clear the marshallers or their instance caches. Looks like your values' object graph is very large, and the cached marshallers are using up a lot of memory that could be garbage-collected (assuming the cache was indeed restarted and only one cache instance is running now). I've created [ISPN-3659] Cache stop should clear thread-local ExtendedRiverMarshaller or their instance caches to address this.

Hi,

Did you find a solution for this? Did you figure out if this is a error in infinspan or a design error?

Im having the exact same problem in my application and im although still using Infinispan Version 5.3.0 final and glassfish 3.1.2.

Any help with this would be great!!