systemml-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Matthias Boehm (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (SYSTEMML-2085) Initial version of local backend
Date Thu, 24 May 2018 18:59:00 GMT

    [ https://issues.apache.org/jira/browse/SYSTEMML-2085?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16489595#comment-16489595
] 

Matthias Boehm commented on SYSTEMML-2085:
------------------------------------------

Well by default, we have a bufferpool, which handles evictions automatically and thus ensures
that you will not run out of memory. The issue you referenced seems, however, related to OpenBLAS
only - apparently they use 2*vcores buffers and fail if you have a total number of threads
larger than that. In order to debug this, please have a look into the assignment of parallelism
to individual operations. If you have 8 vcores, and run the parameter server with a degree
of parallelism of k=4, the operations per worker need to be limited to k=2 - assigning a higher
degree of parallelism would result in performance degradation due to contention. You can handle
this very similar to parfor, which you already extended.  

> Initial version of local backend
> --------------------------------
>
>                 Key: SYSTEMML-2085
>                 URL: https://issues.apache.org/jira/browse/SYSTEMML-2085
>             Project: SystemML
>          Issue Type: Technical task
>            Reporter: Matthias Boehm
>            Assignee: LI Guobao
>            Priority: Major
>
> It aims to implement the local backend for the paramserv function.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

Mime
View raw message