MPerf with an option to sleep for 1ms after processing a message (mimicking a slow receiver) runs out of memory with stock tcp.xml.
Fix: increase MFC.max_block_time from (default) 500ms to 5000ms and/or decrease max_credits.
Investigate why this is the case. I suspect that max_block_time is too low, triggering a request for more credits for all receivers, which respond with the full credits. This replenishes credits at the senders too quickly, not blocking the senders enough, effectively making flow control moot.
Should be fixed with NAKACK4: https://issues.redhat.com/browse/JGRP-2780
- account is impacted by
-
JGRP-2780 NAKACK4: multicast protocol based on positive acks and fixed-size buffer
-
- Resolved
-
