Multi-threading debugging and test

[ZidongLiu]

Summary

• use rng directly in torch.randn instead of relying on the global rng state
• refactor server train one step to make it more readable
• add option in server to switch between 1-by-1 and multi-thread
• test 2 options under 1/5/10 perturb, 1/3/5 local updates, 1/3/5 clients. They should give same result

Result

Before PR:

After PR:

Test

[BichengYing]

Some suggestions about option to switch between serial and parallel and test, you should split that server-client local update out of the main step

[ZidongLiu]

Some suggestions about option to switch between serial and parallel and test, you should split that server-client local update out of the main step

@BichengYing Which step? Any code reference?

[BichengYing]

I mean this def train_one_step(self, iteration: int) -> tuple[float, float]: function. It becomes longer and longer (consider we may add some attack/defend aggregation in the future). It is a good time to split them more. For this PR specifically, this with ThreadPoolExecutor() as executor: part can be extracted out as another sub-function. It will also result in a more reasonable architecture for the switching the parallel and serial I think

[ZidongLiu]

result running test

[ZidongLiu]

main train speed:

this branch parallel=True

this branch parallel=False