search

quiver-team / torch-quiver

PyTorch Library for Low-Latency, High-Throughput Graph Learning on GPUs.
https://torch-quiver.readthedocs.io/en/latest/
Apache License 2.0
293 stars 36 forks source link

A strange phenomenon about quiver sampler #122

Open DesmonDay opened 2 years ago

DesmonDay commented 2 years ago

Hi, thank you for your wonderful opensouce works about Quiver. Recently I was trying to look deep into the sampler module, but met a quite strange phenomenon. We can see the following picture. I was testing about the quiver sampler speed and model training speed.

Machine 1: Tesla V100-SXM2-16GB, Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz, 20 physical CPU cores. Model: GraphSage, batch_size=128, samples=[25,10], Reddit.

屏幕快照 2022-02-28 上午10 45 07 It seems very strange, because the time cost by sampler and model training is less than only using sampler.

I try several methods to figure out the reason, like DataLoader, CUDA Stream, etc, but still confused. Maybe you can help me?

eedalong commented 2 years ago

Well, this is strange

eedalong commented 2 years ago

What is your quiver_sampler's mode

DesmonDay commented 2 years ago

UVA. But I remember that GPU mode has the same phenomenon, if my memory is correct.

eedalong commented 2 years ago

That's odd, let me check it

DesmonDay commented 2 years ago

hi, have you figured out the reason?

LukeLIN-web commented 2 years ago

hi, have you figured out the reason?

Hi, I would like to ask how do you run the quiver. I install torch-quiver 0.1.0 from https://github.com/quiver-team/torch-quiver/blob/main/docker/README.md when i run examples/pyg/reddit_quiver.py, it shows AttributeError: module 'torch_quiver' has no attribute 'device_quiver_from_csr_array'. I tried their docker image and it shows AttributeError: Can't get attribute 'DataEdgeAttr' on <module 'torch_geometric.data.data' from '/opt/conda/lib/python3.7/site-packages/torch_geometric/data/data.py'>