mahout-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Ted Dunning (JIRA)" <j...@apache.org>
Subject [jira] Commented: (MAHOUT-227) Parallel SVM
Date Mon, 21 Dec 2009 05:30:18 GMT

    [ https://issues.apache.org/jira/browse/MAHOUT-227?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12793088#action_12793088
] 

Ted Dunning commented on MAHOUT-227:
------------------------------------

Here are a few formatting suggestions:

a) when cutting and pasting from somebody else's work, it is good to point this out.  You
should directly credit figure 3 and the algorithm pseudo-code which are cut-and-pasted directly
from the original paper.

b) text in your diagram got resized and is now only partially readable.  This makes it a bit
harder to follow exactly what you intend.


More importantly, the parameter k in the original paper is a batch size.  You propose to parallelize
the computation of each batch, but otherwise leave the main structure of the computation in
place.  If we assume a small cluster with, say 100 cores (12 machines or so), then if you
set k to 1000, each core will get to do about a dozen vector operations.  This is likely to
be no more than a microsecond of computation per core per iteration.  My guess is that this
will result in very, very poor CPU utilization since you will require on map-reduce invocation
per iteration.  Concretely put, you will have about a millisecond of useful computation every
10 seconds or so.  

You approach would probably work much better if applied to a single multi-core machine where
the very high rendezvous rate would be more achievable.  I don't expect that this proposed
approach will work with map-reduce.

On the other hand, Pegasos is a pretty scalable algorithm even on a single machine.  If you
were able to produce a high quality sequential implementation, that would be a substantial
contribution to Mahout.


> Parallel SVM
> ------------
>
>                 Key: MAHOUT-227
>                 URL: https://issues.apache.org/jira/browse/MAHOUT-227
>             Project: Mahout
>          Issue Type: Task
>          Components: Classification
>            Reporter: zhao zhendong
>         Attachments: ParallelPegasos.doc, ParallelPegasos.pdf
>
>
> I wrote a proposal of parallel algorithm for SVM training. Any comment is welcome.

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


Mime
View raw message