TensorFlow on Hadoop YARN

[tslam75]

Hadoop YARN is a commonly deployed cluster manager. Having the ability to run TensorFlow on YARN would be very useful in such environment.

Our team is currently working on a YARN application for this purpose, and would like to contribute our work here. We will provide more details of our contribution soon.

-Jason

[jhseu]

Thanks! That'd be really useful.

[zhe-thoughts]

@tslam75 Have you looked at https://issues.apache.org/jira/browse/YARN-6043?

[tslam75]

@zhe-thoughts Thanks for the reference! Looked over YARN-6043, and both uses a native application master for TensorFlow.

Attaching a design document here now. We also have an implementation based on this design, and will publish the code soon.

TensorFlow_on_YARN.pdf

[tslam75]

Sorry for the delay.

Created pull request #39 while waiting for the CLA to be signed.

[IDerr]

Awesome job :O

[leftnoteasy]

In Hadoop 3.0, YARN native services can support running Tensorflow services on YARN without adding any dependencies or implement a new YARN application master.

Please see our blogpost: https://hortonworks.com/blog/distributed-tensorflow-assembly-hadoop-yarn/ and let me know if you have any questions. Thanks!

[tbchj]

focus ...

[butterluo]

@tslam75 Does your 'TensorFlow on Yarn' support fault tolerance ? If yes, how?

[zhanglistar]

mark

[oliverhu]

+1

[zhe-thoughts]

We (LinkedIn Hadoop team) just open sourced TonY: Repo: https://github.com/linkedin/TonY Blog post: https://engineering.linkedin.com/blog/2018/09/open-sourcing-tony--native-support-of-tensorflow-on-hadoop

Comments / discussions very welcome!