systemml-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "LI Guobao (JIRA)" <j...@apache.org>
Subject [jira] [Comment Edited] (SYSTEMML-2396) Batch pre-fetching per workers
Date Thu, 14 Jun 2018 22:19:00 GMT

    [ https://issues.apache.org/jira/browse/SYSTEMML-2396?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16513060#comment-16513060
] 

LI Guobao edited comment on SYSTEMML-2396 at 6/14/18 10:18 PM:
---------------------------------------------------------------

[~mboehm7], is it correct this description? And also, after the last call, I have thought
about the solution that you mentioned which uses the array and slot. However, I could not
figure out what does the solution look like. Does it mean that we launch an another pre-fetch
thread per worker which will produce the sliced matrices and put them into a cache? Or in
each iteration, n more batches will be pre-fetched and saved into a cache so that in the next
few iteration, we can use directly the matrices in this cache?


was (Author: guobao):
[~mboehm7], is it correct this description? And also, after the last call, I have thought
about the solution that you mentioned which uses the array and slot. However, I could not
figure out what does the solution look like. Does it mean that we launch an another pre-fetch
thread per worker which will produce the sliced matrices and put them into a cache?

> Batch pre-fetching per workers
> ------------------------------
>
>                 Key: SYSTEMML-2396
>                 URL: https://issues.apache.org/jira/browse/SYSTEMML-2396
>             Project: SystemML
>          Issue Type: Sub-task
>            Reporter: LI Guobao
>            Assignee: LI Guobao
>            Priority: Major
>
> This task aims to improve the performance of workers. Currently, in each iteration of
mini-batch, we need to slice the matrix, execute the gradients computation and then send them
to the ps for updating the model. While the ps is doing the aggregation work, the worker pauses
due to waiting for the new model. Hence the idea is to completely use this free slot to pre-fetch
the mini-batch in order to accelerate the future iteration.



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

Mime
View raw message