[
https://issues.apache.org/jira/browse/MAHOUT-1573?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14020697#comment-14020697
]
ASF GitHub Bot commented on MAHOUT-1573:
----------------------------------------
Github user dlyubimov commented on the pull request:
https://github.com/apache/mahout/pull/13#issuecomment-45401402
also added (pretty naive) auto parallelism adjustments
(A+B) auto_||
which looks at spark.default.parallelism (P), assumes better parallelism is
acheved if partitions are bumped up to round-ups of 0.95 * P or 1.80 * (P)
whichever is closer. if current partition set is already greater (in
cardinality) than rounded-up 1.80 * (P), all is left as is.
> More explicit parallelism adjustments in math-scala DRM apis; elements of
> automatic re-adjustments
> --------------------------------------------------------------------------------------------------
>
> Key: MAHOUT-1573
> URL: https://issues.apache.org/jira/browse/MAHOUT-1573
> Project: Mahout
> Issue Type: Task
> Affects Versions: 0.9
> Reporter: Dmitriy Lyubimov
> Assignee: Dmitriy Lyubimov
> Fix For: 1.0
>
>
> (1) add minSplit parameter pass-thru to drmFromHDFS to be able to explicitly
> increase parallelism.
> (2) add parrallelism readjustment parameter to a checkpoint() call. This
> implies shuffle-less coalesce() translation to the data set before it is
> requested to be cached (if specified).
> Going forward, we probably should try and figure how we can automate it, at
> least a little bit. For example, the simplest automatic adjustment might
> include re-adjust parallelims on load to simply fit cluster size (95% or 180%
> of cluster size, for example), with some rule-of-thumb safeguards here, e.g.
> we cannot exceed a factor of say 8 (or whatever we configure) in splitting
> each original hdfs split. We should be able to get a reasonable parallelism
> performance out of the box on simple heuristics like that.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
