mahout-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Sebastian Schelter (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (MAHOUT-1431) Comparison of Mahout 0.8 vs mahout 0.9 in EMR
Date Tue, 04 Mar 2014 11:11:21 GMT

    [ https://issues.apache.org/jira/browse/MAHOUT-1431?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13919260#comment-13919260
] 

Sebastian Schelter commented on MAHOUT-1431:
--------------------------------------------

That is really strange, I don't think we changed something in our k-Means implementation (can
someone verify this?). 

Could it be that some change in our vector code causes this behaviour?

Btw: k-Means should have an option to fix the random seed for the initialization to enable
repeatable experiments

> Comparison of Mahout 0.8 vs mahout 0.9 in EMR
> ---------------------------------------------
>
>                 Key: MAHOUT-1431
>                 URL: https://issues.apache.org/jira/browse/MAHOUT-1431
>             Project: Mahout
>          Issue Type: Question
>          Components: Clustering
>    Affects Versions: 0.8, 0.9
>            Reporter: yannis ats
>              Labels: performance
>
> Hi all,
> i tested mahout 0.8 and 0.9 in mahout emr with a large dataset as input and 
> i performed kmeans experiments with both versions in amazon EMR.
> What i found is that mahout 0.8 is faster than mahout 0.9
> in particular i observed that mahout 0.8 is performing less iterations and every iteration
of kmeans is faster than mahout 0.9.Every iteration in mahout 0.8 is twice as fast as that
of 0.9
> the hadoop version was 1.0.x and the input of the data was roughly 2 million datapoints
with dimensionality of 1800.
> The input parameters in both experiments were exactly the same,modulo the initialization
which was random in both cases and i can understand that this may affect the convergence(the
amount of iterations),but i am baffled by the fact that every iteration takes almost twice
the time in 0.9 vs 0.8
> Is this normal?is this  expected?
> thank you in advance for your time.



--
This message was sent by Atlassian JIRA
(v6.2#6252)

Mime
View raw message