Ray on AWS arm Graviton instances


I have been using the c6g instances in AWS based on ARM Graviton processors (with arm Ubuntu 20.04) and found that the performance outperform the Xeon based instances (at least in my python ML training. In fact they have a lower cost than Xeon ones.

The problem is that I want to use Ray on those instances to scale my training to clusters but I found that there is no official ARM based Ray version.

Anyone is in the same boat? Or at least anyone can place me in the right direction on how (if possible) can this be done?

Many thanks in advance, any information is greatly appreciated.


CC @ericl. Any suggestions?

I have built Ray from source in AWS Graviton2 based instances. Refer here for the wheel and some benchmarks between arm vs xeon AWS instances: