tensorflow - Simulate Heterogeneous Clients in TFF

Question

I need to simulate a heterogeneous setting of clients (clients with different hardware resources, e.g different CPU frequencies).

My first thought was to simulate this setting on my local machine by setting different frequencies for each CPU core and then pin each client computation for one of those cores using with tf.device() inside a tff.tf_computation.

The problem is that it seems that we cannot pin a tensorflow computation to a specific core, as discussed in this thread from 5 years ago. I don't know if new functionalities were added regarding this subject.

This answer from another discussion suggests to use distributed tensorflow and then use linux commands to pin specific cores to specific processes. This is the best way to go?

question from:https://stackoverflow.com/questions/66059481/simulate-heterogeneous-clients-in-tff

Answer 1

Waitting for answers