Accurate Communication Time Measurement

[hayekr]

What is your question?

I am working on implementing a federated learning system on real hardware (IoT devices). Currently, I am measuring the communication time by marking the timestamp in the fit method just before the result is returned to the server. However, this timestamp is captured during the fit aggregation phase, which still includes some computation time before the message is serialized and transmitted.

To accurately measure the communication overhead, I need a mechanism to measure the communication time at the exact point of message transmission. Is there any built in way to do this/insert timestamps as a payload in the gRPC packets before transmission?

[WilliamLindskog]

Hi @hayekr

Thanks for asking your question here. There is no direct way of doing this, however it would be a great feature. Are you open to modifying the gRPC layer in Flower to insert custom timestamps, or do you require a solution that works without altering the underlying framework?

[hayekr]

Hi @WilliamLindskog

Thank you for the reply. The goal is to track timestamps as close to the transmission as possible. It would be preferable to do this without modifying the underlying framework, however I am open to modifying the gRPC layer to insert custom timestamps as this would be the most precise solution.

[hayekr]

@WilliamLindskog @jafermarq
It may be possible to copy/route the created_at, delivered_at, pushed_at fields of Task to the transmitted metrics, as opposed to implementing new timestamps.