PL support for GCP instances with Local SSD

[jshin49]

https://cloud.google.com/compute/docs/disks/local-ssd

Google Cloud Platform supports Local SSD, and this supposedly significantly boosts IOPS, which could drastically affect training/testing if there's frequent CPU-GPU communication.

However, I think this may require some Pytorch or Pytorch-lightning level support.

Do you guys have any pointers I can refer to? Or please let me know if I'm wrong.

[kaushikb11]

I think this may require some Pytorch or Pytorch-lightning level support.

Where do you think the support would be required? Have you tried it out with Lightning?

[jshin49]

when I tried using the same code (using lightning) and data on local-ssd enabled GCP instances, there were no notable changes. Meanwhile, this same piece of code would run faster on an on-premise device (which obviously has a dedicated local SSD).