[Q&A] Can I implement asynchronous algorithm on NVFlare? HOW?

[Eric716]

Python version (python3 -V)

3.8

NVFlare version (python3 -m pip list | grep "nvflare")

2.3.0

NVFlare branch (if running examples, please use the branch that corresponds to the NVFlare version, git branch)

No response

Operating system

Ubuntu 20.04

Have you successfully run any of the following examples?

• hello-numpy-sag with simulator
• hello-pt with simulator
• hello-numpy-sag with POC
• hello-pt with POC

Please describe your question

Can I implement asynchronous algorithm on NVFlare?
I want to implement asynchronous algorithm like: FedAsync... on NVFlare.
Does anyone have any suggestions or thoughts? Thank you very much.

[YuanTingHsieh]

@holgerroth can you help comment on this? thanks

[holgerroth]

Thanks for the question @Eric716. Yes, NVFlare can supports asynchronous FL. For example, our controllers provide both blocking and non-blocking versions to send training Tasks to clients, see broadcast and broadcast_and_wait here. The “wait” version is used in synchronous FL workflows, e.g., in ScatterAndGather. Without wait, you could use it to implement asynchronous FL.

Every ‘Task’ that’s been sent to a client has a result_received_cb that can be used that to do some operation on the server side, i.e. add the weight updates to the current global model.

[Eric716]

Thanks! Is there any example for that?

[holgerroth]

We don't have an asynchronous training example at this time. However, you can take a look at this cross-site evaluation workflow to see the asynchronous communication pattern in action. If you end up implementing an example, please consider contributing it back to the repo.