giraph-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Claudio Martella <>
Subject Re: Graph partitioning and data locality
Date Tue, 04 Nov 2014 15:20:21 GMT

answers are inline.

On Tue, Nov 4, 2014 at 8:36 AM, Martin Junghanns <>

> Hi group,
> I got a question concerning the graph partitioning step. If I understood
> the code correctly, the graph is distributed to n partitions by using
> vertexID.hashCode() & n. I got two questions concerning that step.
> 1) Is the whole graph loaded and partitioned only by the Master? This
> would mean, the whole data has to be moved to that Master map job and then
> moved to the physical node the specific worker for the partition runs on.
> As this sounds like a huge overhead, I further inspected the code:
> I saw that there is also a WorkerGraphPartitioner and I assume he calls
> the partitioning method on his local data (lets say his local HDFS blocks)
> and if the resulting partition for a vertex is not himself, the data gets
> moved to that worker, which reduces the overhead. Is this assumption
> correct?

That is correct, workers forward vertex data to the correct worker who is
responsible for that vertex via hash-partitioning (by default), meaning
that the master is not involved.

> 2) Let's say the graph is already partitioned in the file system, e.g.
> blocks on physical nodes contain logical connected graph nodes. Is it
> possible to just read the data as it is and skip the partitioning step? In
> that case I currently assume, that the vertexID should contain the
> partitionID and the custom partitioning would be an identity function in
> that case (instead of hashing or range).

In principle you can. You would need to organize splits so that they
contain all the data for each particular worker, and then assign relevant
splits to the corresponding worker.

> Thanks for your time and help!
> Cheers,
> Martin

   Claudio Martella

View raw message