You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@systemml.apache.org by "Matthias Boehm (JIRA)" <ji...@apache.org> on 2016/08/15 19:11:20 UTC

[jira] [Closed] (SYSTEMML-809) Performance issue multi-threaded wdivmm_left

     [ https://issues.apache.org/jira/browse/SYSTEMML-809?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Matthias Boehm closed SYSTEMML-809.
-----------------------------------

> Performance issue multi-threaded wdivmm_left
> --------------------------------------------
>
>                 Key: SYSTEMML-809
>                 URL: https://issues.apache.org/jira/browse/SYSTEMML-809
>             Project: SystemML
>          Issue Type: Bug
>          Components: Runtime
>            Reporter: Matthias Boehm
>            Assignee: Matthias Boehm
>             Fix For: SystemML 0.11
>
>
> Recent experiments of wdivmm left (e.g., t(U) %*% (X/(U%*%t(V)+eps))) and right (e.g., (X/(U%*%t(V)+eps))%*%V) revealed severe performance issues of wdivmm left with increasing sparsity. While wdivmm right shows good multi-threading speedups of 9x-10x (with 12 physical cores), wdivmm left shows only a speedup of 3x.
> Cause: After detailed analysis, the cause was a high ratio of L3 cache misses on the output despite the existing cache-conscious blocking and column-block-wise parallelization. This is due to the row major sparse input X, where an input row Xi touches a single output row for wdivmm right, but potentially all output rows for wdivmm left.  



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)