Details
-
Bug
-
Resolution: Unresolved
-
Minor
-
None
-
None
-
3
-
9223372036854775807
Description
There are scenarios which result in short IO, which means a lot of "immediate" messages are being sent.
If node is receiving a lot of "immediate" messages without having to send anything back, peer credits may be returned to the sender too late, causing the sender to run out of credits.
"Immediate" message in o2iblnd terms is a short, less than 4K, message which doesn't require an ack/nack to be sent back.
Currently when immediate message is received, the recipient is not checking whether hiw (high water mark) level is exceeded and it is time to return the credits (by sending a noop message). If the node is receiving only the immediate messages for a while, then only the periodic "check_connections" routine is checking whether hiw mark is exceeded, which may not be quick enough to avoid unnecessary delays.