Tutorial on boosting pina performance (speed and accuracy)

[dario-coscia]

[njwangcn]

I feel that the data processing part of Pina seems to limit the performance of GPU acceleration. In other words, the data processing cannot keep up with the speed of GPU computation. Does anyone have good solutions for this?

[dario-coscia]

@njwangcn Yes we are indeed aware of the issue. In the new version 0.2 which we plan to release very soon we have added a completely new datamodule. For the user nothing will change on front end, but we will be way faster on GPU and multiGPUs training will be added!

[njwangcn]

Thank you for your response! I really appreciate the hard work and dedication your team has put into improving the Pina project. It’s exciting to hear about the new datamodule and the performance enhancements coming in version 0.2.

[dario-coscia]

Hi @njwangcn ! I will close this issue, thanks for you feedback :-). We are adding many new tutorials in #544, with a dedicated tutorial on PINA Trainer to show how to get max accuracy and performance.

PS.
We have added #423 different options to speed up the data processing part in PINA. Now, if you datat batch fits in GP,U we move it only once, and you gain roughly a 50 % speedup compared to 0.1 version. Also, you can now use multiple GPUs for Distributed Data training if you have a lot of data, decreasing the computational time. Finally, we added for PINN training fast differential operators #528 which can be used instead of the already presented ones, at the cost of not having internal checks (I suggest first to use the slow ones, and once everything work just change in the import statement to the fast version)