search

K-Wu / pytorch-direct_dgl

PyTorch-Direct code on top of PyTorch-1.8.0nightly (e152ca5) for Large Graph Convolutional Network Training with GPU-Oriented Data Communication Architecture (accepted by PVLDB)
https://arxiv.org/abs/2103.03330
45 stars 5 forks source link

The role of zero-copy #9

Open Changyuan0825 opened 6 months ago

Changyuan0825 commented 6 months ago

Dear Kun Wu,

I hope this message finds you well.

I have a few questions regarding your paper, "Large Graph Convolutional Network Training with GPU-Oriented Data Communication Architecture." When the GPU starts training the GNN, is all the data already copied onto the GPU? Is the use of zero-copy technology intended to disperse data reading from the CPU to the GPU to avoid the traditional DMA's need to perform gather operations? In other words, is zero-copy used to manage the transfer of data to the GPU more efficiently, rather than having the GNN operations directly read from the CPU DRAM during training?

Thank you for your time and clarification.

Best regards, Changyuan

K-Wu commented 5 months ago

Hi Changyuan,

Thank you for your interest in our work! You are correct in all three of your questions. Please let me know if you have further questions after reading our paper and using our code.

I believe you have watched it, but just FYI, the GTC presentation (https://www.nvidia.com/en-us/on-demand/session/gtcspring21-s32038/) is a good resource for quickly learning about our work and establishing intuition.

Best Regards, Kun

Changyuan0825 commented 4 months ago

Thank you for your help!

With your guidance, I have come to appreciate the brilliance of your work.

Best regards, Changyuan