On Wed, Jun 20, 2012 at 6:19 AM, Sebastian Schelter <[email protected]> wrote:

> I'm just asking because I review a paper where Mahout's SVM support on
> M/R is mentioned :)
>

Interesting.  We don't have anything release, I don't think.


> As far as I recall All-Reduce is map + an aggregation tree that feeds
> back the result?
>

Yes.  Exactly, except that there isn't a map step.  It typically also
requires the use of combiners so that it can use a tree to broadcast to the
aggregating node.  This allows O(n / log n) average parallelism during
aggregation and is critical for performance.


> Giraph supports aggregators where every worker instance preaggregates
> the data, the master computes the final aggregation and feeds it back to
> the workers. So you could have a 1-level aggregation tree out of the box.
>

That won't cut it in the end, but might be useful in the short term since
it avoids the horrible map-reduce iteration cost.  The goal is to sum a
(big) vector from each node.  A 1-level aggregation tree requires that all
the vectors be copied to a single node where the sums will be done.


> Best,
> Sebastian
>
> On 20.06.2012 15:07, Ted Dunning wrote:
> > No.
> >
> > But L1 regularization for hinge loss should give essentially identical
> results. We currently support logit loss but this is pluggable.  This is in
> the sgd framework which currently lacks a map reduce implementation.
> >
> > Speaking of that, does anybody offhand know whether giraph supports an
> allreduce primitive. That is the primary thing missing to get parallel sgd.
> >
> > Sent from my iPhone
> >
> > On Jun 20, 2012, at 6:01 AM, Sebastian Schelter <[email protected]> wrote:
> >
> >> Hi,
> >>
> >> Did we ever have a M/R implementation of an SVM?
> >>
> >> Best,
> >> Sebastian
>
>

Reply via email to