The caches in the 'infinispan' subsystem configured with <transaction mode="NONE"/> (which results in no transaction used by 2LC) get stuck in state= CLEARING when you perform a programmatic evict() on them to clear the caches. (Using e.g. em.getEntityManagerFactory().getCache().evictAll() )
The problem seem to lie within the class 'org.hibernate.cache.infinispan.impl.BaseRegion, where the 'checkValid' method fails to clear the cache (and set back the state to VALID). The reason it fails is due to that it calls: cacheAdapter.withinTx() , which in turn tries to get the TransactionManager from the cache (which seems to be null for these non-transactional caches), and then NPEs on the tm.begin(); The NPE is caught and logs out (TRACE in org.hibernate.cache.infinispan.impl.BaseRegion) "Could not invalidate region: null", and never set the state in VALID. So from thereon the caches are in an Invalid state and is not used anymore by Infinispan, thus leading to much performance.