The cache used for clustered SFSBs uses FileCacheLoader, which creates one directory per passivated bean instance. This means the max number of passivated beans is affected by the filesystem's max number of child directories. One of the causes of the issues reported in JBPAPP-90 is a load test is hitting a limit of ~ 64K nodes on a 64 bit ext3 filesystem. Some google hits related to this make me think in some cases the limit might be 32K.

The default removal timeout for ejb3 SFSBs is 0 (i.e. no removal), so hitting this limit is not far-fetched.

Some possible solutions:

1) Add another layer to the Fqn, hash the bean id, and spread the beans among buckets.
2) Use another CacheLoader.