Can I run smallpond with a remote ray cluster?

[JustinDuy]

Hi everyone, it is very impressive when one looks at the throughput. I just wonder if i can integrate smallpond usage with a RAY_ADDRESS on remote k8s ray cluster, so each remote ray task will be executed on the cloud and not my local machine? Maybe i overlook the documentation, but if anyone could point me to it, i would appreciate.

[gl2007]

Hi everyone, it is very impressive when one looks at the throughput. I just wonder if i can integrate smallpond usage with a RAY_ADDRESS on remote k8s ray cluster, so each remote ray task will be executed on the cloud and not my local machine? Maybe i overlook the documentation, but if anyone could point me to it, i would appreciate.

Might be worthwhile to execute the whole thing in the cloud; your throughout might suffer if you mix on premise and load the "sets" into cloud on the fly, just saying, not that I am an expert. But I did have the opportunity to load huge datasets via ADF and process in cloud and that was my experience.

[JustinDuy]

you are right. I am thinking of all 'on-premise', both compute and storage.

[wangrunji0408]

Hi, theoretically, it should be possible. However, we haven't tested this specific setup. As long as both of your local machine and the remote Ray cluster can access a high-performance shared storage, it should work.

[JustinDuy]

yeah i think so too, in my imagination, a high-performance NFS PVC served as shared storage should work but i couldn't quite figure out how to do it with 3fs...

[junli1026]

It will be fun to see Ray running on 3FS, whose shared object spilling mechanism can benefit a lot from 3FS high performce.