You are viewing a plain text version of this content. The canonical link for it is here.

Posted to dev@hama.apache.org by Suraj Menon <me...@gmail.com> on 2012/11/12 19:04:15 UTC

Question on GraphJobRunner

In the GraphJobRunner implementation, during the runtime-partitioning
phase, each peer reads a vertex from input and sends the vertex to the
 peer that holds the partition and then gets into sync.  Why are we not
reading all the vertices and sending all vertices to their respective peers
all at once in one sync? Did we hit a scalability bottleneck here?

Just found this when I was re-factoring the graph code with Superstep
chaining. Apologies, if this was already discussed and I am late to the
party.

-Suraj

Re: Question on GraphJobRunner

Posted by Thomas Jungblut <th...@gmail.com>.

>
> Did we hit a scalability bottleneck here?


yes.

2012/11/12 Suraj Menon <me...@gmail.com>

> In the GraphJobRunner implementation, during the runtime-partitioning
> phase, each peer reads a vertex from input and sends the vertex to the
>  peer that holds the partition and then gets into sync.  Why are we not
> reading all the vertices and sending all vertices to their respective peers
> all at once in one sync? Did we hit a scalability bottleneck here?
>
> Just found this when I was re-factoring the graph code with Superstep
> chaining. Apologies, if this was already discussed and I am late to the
> party.
>
> -Suraj
>