The data passthrough path from pm to GPU

[pyhundan]

Hi, I read your article GPM: I have also looked at some of your source code, but there are some things I still don't understand. I want to use an example code to let me know the workflow of GPM, but I can't find the corresponding code in this repository. Can you upload a sample code on how to call the interface of the GPM library?

Thanks, hundan

[AKKamath]

Hi Hundan, Thanks for the feedback. The source files in the GPMBench folder call the library interface, for example DNN (https://github.com/csl-iisc/GPM-ASPLOS22/blob/master/GPMBench/checkpointing/DNN/src/lenet.cu). If you search for GPM on that page, you will see the library calls.

Within a few days, I'll also add a few small example programs demonstrating the different function calls of the library. Hopefully that'll make things a bit more clear.

Let me know if anything else is needed.

[pyhundan]

Hi, @AKKamath Thanks for your reply. I am very interested in the process of transferring data from PM to GPU, and I hope you can provide some simple examples to demonstrate how to use GPM lib API to transfer data from PM file to GPU and persist data from GPU to PM, like a simple file not a network dataset.

I have read the example DNN. I have a question. According to the paper, data on the GPU can be written directly to the PM. The interface called is gpmcp_restore () ? such as line 1020, at other times I don't see data written to a file, except for the last save file. But that's where the comments describe writing to the CPU.

Thanks a lot.

[pyhundan]

Hi, @AKKamath I read about a microbenchmark in Figure 3 in my paper. Can I get the code for this microbenchmark?

Thanks a lot.