incubator-giraph-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Avery Ching <ach...@apache.org>
Subject Re: Did GIRAPH-11 break vertex reactivation?
Date Tue, 15 Nov 2011 21:12:56 GMT
This should fix it.  It passed local unittests.  Let me know.

Avery

On 11/15/11 1:03 PM, Avery Ching wrote:
> Yes, I think I broke it.  Sorry.  Let me get you a diff to test quickly.
>
> Avery
>
> On 11/15/11 12:42 PM, Sebastian Schelter wrote:
>> Hi,
>>
>> I updated to the latest trunk (after the GIRAPH-11 commit) and wanted to
>> continue to work on GIRAPH-51 where I use a small toy graph to test
>> SimpleShortestPathVertex.
>>
>> Unfortunately my code did not work anymore and I guess I tracked it down
>> to the fact that vertex that voted to halt are not reacted anymore when
>> new messages arrive.
>>
>> In SimpleShortestPathVertex every vertex always votes to halt and only
>> gets reactivated when a shorter path to it has been found. However my
>> test run always finished after superstep 0.
>>
>> I don't know too much about Giraph's internals yet, but my guess is that
>> the number of sent messages is not tracked correctly anymore. Therefore
>> giraph finishes the algorithm (as all vertices voted to halt) although
>> there should still be messages in the pipeline.
>>
>> I think I tracked it down to this behavior:
>>
>> GraphMapper declares a variable workerSentMessages = 0 and never
>> increases it. This variable is given to
>> BspServiceWorker.finishSuperstep() which writes it to zookeeper and uses
>> it to compute the GlobalStats afterwards, which are used to decide
>> whether a new superstep has to be scheduled. As it has never been
>> increased, the algorithm will always stop when all vertices voted to 
>> halt.
>>
>> It would be great if someone could confirm/disprove this speculation and
>> help me to continue work on GIRAPH-51
>>
>> --sebastian
>


Mime
View raw message