giraph-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Alessandro Presta <>
Subject Re: Review Request: Out-of-core messages
Date Sun, 22 Jul 2012 20:49:46 GMT
Exactly. On paper, the amount of data around should be the same as during
the computation, but in practice we do use a lot more memory.
You can play with the settings and just push the problem a little farther
away, by caching less and flushing requests more frequently, so then the
bottleneck is on the servers.
We're basically sending (k-1)/k of the graph through the network, where k
is the number of workers.

What I'm thinking is that in INPUT_SUPERSTEP we're doing what MapReduce is
really good at (sorting and aggregating) in a probably inefficient (or at
least non-scalable) way.
We could try implementing it with a MapReduce job instead, where the
mappers take input splits and emit (partition_id, vertex) (they would have
access to the partitioner) and reducers just output the built partitions
to HDFS.
The computation stage would then be the usual Giraph job, where each
worker knows where to get its partitions from HDFS.
I can try making this change and see how it goes. It would just be one MR
job, so we're not selling our souls to iterative MR.

I can also see many cases where one might not want to shuffle vertices
around at all: each worker reads a roughly equal part of the input (forget
about bigger vertices for now) and simply communicates its own vertex ids
to the master. Partition "a posteriori" instead of "a priori".

What do you think?

On 7/20/12 9:42 PM, "Eli Reisman" <> wrote:

>What we are seeing in the metrics is the three-way load of
>1. reading InputSplits from HDFS (mostly over the wire as there is no
>locality right now)
>2. creating temporary collections of vertices, sending them on netty
>3. simultaneously receiving collections of vertices on netty from remote
>nodes that will be place in the local workers' partitions for processing

View raw message