By any chance using broadcast queue?
check CSCsr17044
Try to use the PUB, first
Then in the AC machines configure hosts files
pointing to both SUB and PUB IP addresses.
If same issue I may open a TAC case and upload CCM SDI, SDL, CTI SDI SDL and AC detailed traces 15 minutes before the failure and 15 after the service restarted.
Also Syslog and Perfmon logs