Test and add a description of distributed computation about deepwave

[vkazei]

hi Alan,
Could you share some thoughts on the easiest way to distribute shots?
Best,
Vladimir

[ar4]

Hi Vladimir, PyTorch provides several options for distributed execution: https://pytorch.org/tutorials/beginner/dist_overview.html . It should also be possible to achieve it manually using mpi4py, if you prefer. Do you wish to run on multiple GPUs connected to a single node, multiple nodes each with a single device (CPU or GPU), or multiple nodes each with multiple GPUs? Are you just doing forward modeling, or also backpropagation/inversion?

…

-Alan

[vkazei]

Thanks a lot for getting back. I am trying to distribute FWI and LSRTM within a single node. I tried to replace prop=nn.DataParallel(prop), which appears to be the most basic option.
source_amplitudes.swapaxes(0,1).swapaxes(1,2) before the propagator and swapping back within the propagator allows distributing inputs seemingly properly. The model itself does not get replicated staying at the "cuda:0" device so even forward propagation does not work...

[ar4]

Hi Vladimir,

I think I see the problem. The model is not being registered as a parameter of the propagator, and so PyTorch doesn't know that it needs to copy it to the other devices. I don't have multiple GPUs to test it on at the moment, but does manually registering the parameters as below work?

prop = deepwave.scalar.Propagator({'vp': model}, dx)
prop.register_parameter('vp', torch.nn.Parameter(model))
prop = torch.nn.DataParallel(prop)

If so, I will fix it in the next release of Deepwave so that the parameters are automatically registered.

[vkazei]

Hi Alan,

Registering the model as a parameter did not change the behavior. Sending it manually to the device with inputs inside the forward method lets nn.DataParallel run, but it looks like it runs sequentially the forward propagation for different GPUs)

[ar4]

Hi Vladimir, That is unfortunate. It should definitely be possible, though. Perhaps larger changes are required. I planned to overhaul Deepwave during the summer, but will try to get to it sooner. Please let me know if you can think of any other features that you would like it to have so that I can try to include them when planning the modifications. In the meantime, another option, if you are very keen to use multiple GPUs and feeling ambitious, might be to use mpi4py. It will be more work, as you will need to manually use mpi4py to run multiple processes and send data and gradient updates between them, but should provide you with enough control that you can make it work. If you can wait a few weeks, though, I will hopefully have a clean way working.