If Infinispan DIST is used, at least one of the cache locations for an SFSB should be the node that created it. This way in the normal case where there are 2 copies, a single unicast to the backup node is sufficient to replicate. If the cache key (the SFSBContext id) hashes to 2 remote nodes, then every cache write will require 2 unicasts.

The problem here is unlike web sessions, the id for an SFSB context is not generated by the caching layer; it's generated by StatefulBeanContext itself. So the web session strategy of having a background thread occasionally populate a pool of ids that are known to map locally (using the consistent hash algorithm Mircea Marcus is going to expose for ISPN-232) won't work.