[
https://issues.apache.org/jira/browse/MADLIB-1428?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Domino Valdano updated MADLIB-1428:
-----------------------------------
This has been merged. Idle time between hops for VGG-16 has been reduced from
50s down to less than 10s. Combined with new caching option, JIT auto
clustering (enabled now by default), and reduced initialization time, v1.18
should be much faster.
> DL - reduce GPU idle time between hops
> --------------------------------------
>
> Key: MADLIB-1428
> URL: https://issues.apache.org/jira/browse/MADLIB-1428
> Project: Apache MADlib
> Issue Type: Improvement
> Components: Deep Learning
> Affects Versions: v1.17
> Reporter: Frank McQuillan
> Assignee: Domino Valdano
> Priority: Major
> Fix For: v1.18.0
>
>
> For large models like VGG16, we are seeing ~50 sec idle GPU time while model
> state is hopped between segments. We should investigate ways to reduce this
> idle GPU time.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
