[
https://issues.apache.org/jira/browse/MAHOUT-922?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13167329#comment-13167329
]
Dmitriy Lyubimov commented on MAHOUT-922:
-----------------------------------------
and oh yeah i use double[][] for blocks to ensure proper allocation and
otherwise simplification. DenseMatrix contract does not guarantee specific
memory allocations contracts for skinny vs. wide matrices that affect efficient
memory use in this case greatly.
> SSVD: ABt Job tweaks for extra sparse inputs
> --------------------------------------------
>
> Key: MAHOUT-922
> URL: https://issues.apache.org/jira/browse/MAHOUT-922
> Project: Mahout
> Issue Type: Improvement
> Components: Math
> Affects Versions: 0.6
> Reporter: Dmitriy Lyubimov
> Assignee: Dmitriy Lyubimov
> Fix For: 0.6
>
> Attachments: MAHOUT-922.patch
>
>
> Per tests on Sebastian's extremely sparse large inputs (4.5m x 4.5 m).
> AB' performance is still a bottleneck if one uses power iterations. For
> sufficiently sparse inputs it may turn out that mappers cannot form the
> entire blocked product in memory for Y_i. the Y_i block is going to be of
> size s x (k+p) where s is number of A rows read in a given split. in cases
> when A is extra sparse, such blocks may actually take more space than the A
> input. When this happens, s is constrained by -oh parameter and combiners and
> reducers get flooded by partial oh x (k+p) outer products and seem to have
> hard time to sort and shuffle them (especially high pressure on combiners has
> been seen).
> So, several improvements in this patch:
> -- present Y_i blocks as dense (they are beleived to be dense anyway, so
> keeping them as sparse just eats up RAM by sparse encoding, so at least twice
> as high blocks can actually be formed);
> -- eliminate combining completely. instead of persisting and sorting and
> summing up partial product in combiner, sum up map-side. if block height is
> still insufficient and cannot be extended due RAM constraints (unlikely for
> Sebastien's 4.5 x 4.5 mln case) just perform additional passes over B'. Since
> computation is cpu bound, additional passes over B' should not register.
> However, elimination of combiner phase for high load cases is probably going
> to have a dramatic effect.
> -- set max block height for Q'A and AB' separately instead of single -oh
> option. Their scaling seems to be quite different in terms of OOM danger. in
> my experiments Q'A blocking enters red zone at ~150,000 already whereas AB'
> block height can freely roam over a million easily for the same RAM. I
> provide 200,000 (~160Mb for k+p=100) as a default for AB' blocks which should
> be enough for Sebastien's 4.5 x 4.5 mln sparse case without causing more than
> one block.
> Miscellanea:
> -- Test run time: removed redundant tests and checks for SSVD. reduced test
> input size.
> -- Per Nathan's suggestion, p parameter is now optional, default is 15
> (computation time is cubic to it, so I want to be careful not to run it too
> high by default).
> Current patch branch work is here:
> https://github.com/dlyubimov/mahout-commits/tree/MAHOUT-922
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
