You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@spark.apache.org by "WuZeyi (Jira)" <ji...@apache.org> on 2019/11/19 08:27:00 UTC

[jira] [Comment Edited] (SPARK-29553) This problemis about using native BLAS to improvement ML/MLLIB performance

    [ https://issues.apache.org/jira/browse/SPARK-29553?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16977236#comment-16977236 ] 

WuZeyi edited comment on SPARK-29553 at 11/19/19 8:26 AM:
----------------------------------------------------------

Thanks for Looking.

I modify spark-env.sh,and the flame graph is like this:

!image-2019-11-19-16-11-43-130.png!

Then I modify {color:#172b4d}spark.conf, and the flame graph is this:{color}

{color:#172b4d}!image-2019-11-19-16-13-30-723.png!{color}

It means that it is still multi-thread if I modify spark-env.sh.

If I modify spark.conf to set  {color:#ff0000}spark.executorEnv.OPENBLAS_NUM_THREADS=1{color},it works and the the performance improve.

IMHO, spark-env.sh only set the evn of the spark-sumbit process, but it doesn't work in the executor processes.


was (Author: zeyiii):
Thanks for Looking.

I modify spark-env.sh,and the flame graph is like this:

!image-2019-11-19-16-11-43-130.png!

Then I modify {color:#172b4d}spark.conf, and the flame graph is this:{color}

{color:#172b4d}!image-2019-11-19-16-13-30-723.png!{color}

It means that it is still multi-thread if I modify spark-env.sh.

If I modify spark.conf to set  {color:#ff0000}spark.executorEnv.OPENBLAS_NUM_THREADS=1{color},it works and the the performance improve.

IMHO, spark-env,sh only set the evn of the spark-sumbit process, but it doesn't work in the executor processes.

> This problemis about using native BLAS to improvement ML/MLLIB performance
> --------------------------------------------------------------------------
>
>                 Key: SPARK-29553
>                 URL: https://issues.apache.org/jira/browse/SPARK-29553
>             Project: Spark
>          Issue Type: Improvement
>          Components: ML, MLlib
>    Affects Versions: 2.3.0, 2.4.4
>            Reporter: WuZeyi
>            Priority: Minor
>              Labels: performance
>         Attachments: image-2019-11-19-16-11-43-130.png, image-2019-11-19-16-13-30-723.png
>
>
> I use {color:#ff0000}native BLAS{color} to improvement ML/MLLIB performance on Yarn.
> The file {color:#ff0000}spark-env.sh{color} which is modified by SPARK-21305 said that I should set {color:#ff0000}OPENBLAS_NUM_THREADS=1{color} to disable multi-threading of OpenBLAS, but it does not take effect.
> I modify {color:#ff0000}spark.conf{color} to set  {color:#FF0000}spark.executorEnv.OPENBLAS_NUM_THREADS=1{color},and the performance improve.
>   
>   
>  I think MKL_NUM_THREADS is the same.
>   



--
This message was sent by Atlassian Jira
(v8.3.4#803005)

---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@spark.apache.org
For additional commands, e-mail: issues-help@spark.apache.org