Bug fix in sendMessage() of pregel implementation in PageRank.scala

[mangolzy]

What changes were proposed in this pull request?

Iterator((edge.dstId, edge.srcAttr._2 * edge.attr))
->
Iterator((edge.dstId, edge.srcAttr._1 * edge.attr))

How was this patch tested?

Since edge.srcAttr._2 is used to compare with tol, it should be the (newPR - oldPR), but in the sendMessage, the origin code send it as part of the message, instead it should be the newPR which is edge.srcAttr._1.

Please review http://spark.apache.org/contributing.html before opening a pull request.

[WenqianZhao]

You can check the comment in pull request #2100.

[WenqianZhao]

By the way, @ankurdave , hi, I'm very curious about why you use delta instead of just newPR as the message. Is that for better performance or just at chance? Could you please answer this question when you are free? Thank you so much.

[ankurdave]

@WenqianZhao I think the point of sending deltas instead of absolute ranks was that, as parts of the graph converge, their deltas would go to zero. GraphX would then be able to compress those zero messages more efficiently.

[WenqianZhao]

@WenqianZhao I think the point of sending deltas instead of absolute ranks was that, as parts of the graph converge, their deltas would go to zero. GraphX would then be able to compress those zero messages more efficiently.

Got it! Thank you!

[AmplabJenkins]

Can one of the admins verify this patch?