Support gang-scheduling by kube-batch

Question

Support gang-scheduling by kube-batch

k82cn opened this issue 6 years ago · comments

Gang-scheduling is a common requirement from training job; kube-batch supports it right now :) So open this issue to trace the discussion.

Klaus Ma · Answer 1 · Tue Jan 29 2019 14:59:17 GMT+0800 (China Standard Time)

/kind feature

Klaus Ma · Answer 2 · Thu Feb 14 2019 19:30:14 GMT+0800 (China Standard Time)

@johnugeorge , what's the plan of 0.5?

Johnu George · Answer 3 · Fri Feb 15 2019 17:52:53 GMT+0800 (China Standard Time)

@k82cn
Currently, gang scheduling behavior is consistent across TF and Pytorch operators.

Are you working on kubeflow/training-operator#916? This is planned for 0.5 release across operators.

Klaus Ma · Answer 4 · Wed Mar 27 2019 09:29:19 GMT+0800 (China Standard Time)

regarding kubeflow/training-operator#916 , replace PDB with PodGroup is done; I'm thinking how to add some other advanced feature from kube-batch by PodGroup :). I'll do some investigation after 0.5 :)