i have a flask api tunning a pytorch predict inside a kubernetes cluster. I deployed it on AWS and got good results.

I have tried on GKE and had good results, but when im deploying my pods on a high cpu node that has over 32 vcpu, the predict take from 30ms to 1300 ms.

Deploying my pod on a node that has less cpu makes it run at 30ms.

It makes no sense for me and i'm unable to get how's that possible.

  • 1
    What type of compute nodes and storage are you using on each platform?
    – casey vega
    Aug 2 '19 at 16:10
  • Could you reply to the comment?
    – 030
    Dec 25 '19 at 11:52