From an email from Bela:
OK, when running async DIST with 2 nodes and numOwners == 2, I found a huge memory leak in the L1ManagerImpl !
With L1 enabled, my test with 1 passive and 1 active node wouldn't complete with 10 threads each putting 100'000 elements in the cache; instead the passive node would always OOME !
With L1 disabled, it would complete, with 23'000 TXs / sec !
Profiling the passive node (the one which would receive the modifications via RPCs), I can see that L1ManageImpl.addRequestor() leaks entries in the 'requestors' hashmap.
The weird thing is that L1 shouldn't even matter in the 2 node case !