[jboss-jira] [JBoss JIRA] (JGRP-1675) Threads stuck in FlowControl.decrementIfEnoughCredits

[Bela Ban (JIRA)]

    [ https://issues.jboss.org/browse/JGRP-1675?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12850944#comment-12850944 ] 

Bela Ban commented on JGRP-1675:
--------------------------------

UPerf2's GET request uses a GET_FIRST on 2 members. This clogs up the receivers' OOB thread pools, and thus credits are dropped. Sending the CREDIT_REQUESTS as OOB fixes this, but the performance is still slow because of the exhaused OOB pool.

I suggest send the GET as a unicast to the owner and if the owner doesn't respond within N ms (perhaps computed as 30% over the rolling avg time for GETs), then try the other owner (essentially staggered GETs as suggested by Manik).

Alternatively, a GET could be an anycast, but the target list should include only the primary owner, conceptually this is the same as above.
                
> Threads stuck in FlowControl.decrementIfEnoughCredits
> -----------------------------------------------------
>
>                 Key: JGRP-1675
>                 URL: https://issues.jboss.org/browse/JGRP-1675
>             Project: JGroups
>          Issue Type: Bug
>    Affects Versions: 3.4
>            Reporter: Radim Vansa
>            Assignee: Bela Ban
>             Fix For: 3.5
>
>         Attachments: jgroups-udp-radim.xml, RemoteGetStressTest.java, UPerf2.java
>
>
> I have recently observed a repeated situation where many (or all) threads have been stuck waiting for credits in FlowControl protocol.
> The credit request was not handled on the other node as this is non-oob message and some (actually many of them - cause unknown) messages before the request have been lost - therefore the request was waiting for them to be re-sent.
> However, these have not been re-sent properly as the retransmission request was not received - all OOB threads were stuck in the FlowControl protocol as these handled some other request and tried to send a response - but the response could not be sent until FlowControl gets the credits.
> The probability of such situation could be lowered by tagging the credit request to be OOB - then it would be handled immediately. If the credit replenish message would then be processed in regular OOB pool, this could get already depleted by many requests, but setting up the internal thread pool would solve the problem.
> Other consideration would be to allow releasing thread from FlowControl (let it send the message even without credits) if it waits there for too long.
> h3. Workaround
> It appears that setting MFC and UFC max_credits to 10M or removing these protocols at all is a workaround for this issue.

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira