Adam,

Have you tested with a downsampled user history matrix?

That could have massive effect on run-time.

A parallelized version should benefit similarly.



On Mon, Mar 3, 2014 at 11:50 AM, Adam Ilardi (JIRA) <[email protected]> wrote:

>
>     [
> https://issues.apache.org/jira/browse/MAHOUT-1429?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13918358#comment-13918358]
>
> Adam Ilardi edited comment on MAHOUT-1429 at 3/3/14 7:48 PM:
> -------------------------------------------------------------
>
> Nothing formal. Anecdotally my total runtime has gone way down 2-3 days to
> < one day. I noticed that it would be stuck on one core for a very long
> time which led me to this code. I run the model with ~9 million items.
>
>
> was (Author: adamilardi):
> Nothing formal. Anecdotally my total runtime has gone way done 2-3 days to
> < one day. I noticed that it would be stuck on one core for a very long
> time which led me to this code. I run the model with ~9 million items.
>
> > Parallelize YtransposeY in ImplicitFeedbackAlternatingLeastSquaresSolver
> > ------------------------------------------------------------------------
> >
> >                 Key: MAHOUT-1429
> >                 URL: https://issues.apache.org/jira/browse/MAHOUT-1429
> >             Project: Mahout
> >          Issue Type: Improvement
> >          Components: Collaborative Filtering
> >    Affects Versions: 0.9
> >            Reporter: Adam Ilardi
> >            Priority: Minor
> >              Labels: newbie, patch, performance
> >         Attachments: MAHOUT-1429.patch
> >
> >
> > I wrote a simple patch to do the calculation in parallel. It could be
> better but it gets the job done. Is there some other (Matrix transpose dot
> matrix) code in mahout that is already parallel? I will reuse that if it
> exists.
>
>
>
> --
> This message was sent by Atlassian JIRA
> (v6.2#6252)
>

Reply via email to