[A Job has tasks running in parallel each contained in a peer.] Every job has its own synchronization barrier. If one of your task is stuck on infinite loop, it will hold up other tasks in sync. Other jobs have their own sync barrier. Hence the whole cluster would not be blocked.
>> And are the messages really only sent over the network when a sync is reached? This would cause large spikes in network traffic. I would imagine it would be more network friendly to send the messages asap, but buffer them on the receiving end, until the sync is reached, thereby reducing the network traffic spike on a sync to only the sync signal itself. https://issues.apache.org/jira/browse/HAMA-490 issue is for implement asynchronous communication. >> Can you have different groups of jobs that need their own independent sync cycles, but still want to communicate with other groups? Do you need to use an external API to do the communication in that case? https://issues.apache.org/jira/browse/HAMA-639 issue would have a provision to define tasks in this way. However, the message queue design is not thought out for this. -Suraj On Sat, Sep 22, 2012 at 9:36 AM, Diot Sebastien <[email protected]> wrote: > Hi. > > What happens if one job "get stuck", and doesn't sync anymore? Does the > whole Hama cluster grind to an halt? > > Can you have different groups of jobs that need their own independent sync > cycles, but still want to communicate with other groups? Do you need to use > an external API to do the communication in that case? > > > > MfG / Regards, > Sebastien Diot > > Grossblittersdorferstrasse 257-259 > 66119 Saarbruecken > Germany > Tel.: +49 (681) 8808-0 > Fax.: +49 (681) 8808-300 > E-Mail.: [email protected] > ----------------------------------------------------- > EURO DATA GmbH & Co. KG > Datenverarbeitungsdienst > HR A 6448 Amtsgericht Saarbrücken > Komplementär: A. Reiß & Sohn GmbH > HR B 4965 Amtsgericht Saarbrücken > Geschäftsführer: Dipl.-Kfm. Karl-Heinz Siebenpfeiffer > ----------------------------------------------------- > > >
