Is the code for GRAPH PARTITIONER included in this repo?

[molamooo]

The paper of ROC introduces a learned-based graph partitioner, but there seems to be no related code included in this repo?

[happy233cat]

Hi, Jeffery

did you find the code for learning-based graph partitioner? I am also looking for it

Thanks!

[molamooo]

Sorry I didn't find related codes. But I found this self hosted dataset. Not sure if it's preprocessed by the learning-based partitioner. link

[happy233cat]

Got it, That's super helpful!
did you manage to compile the whole project to generate the gnn executable?
I was encountering some compilation errors

[molamooo]

I was able to compile and run the project with no issue.

[happy233cat]

which version of cuda you are using? I am currently using CUDA 11.0 for ampere GPU

[molamooo]

It's been a long time... I think I run it using CUDA 10.2/10.1 for volta GPU.

[happy233cat]

Got it, I will try with CUDA 10.1/10.2, there are some issue with compilation on CUDA-11

[happy233cat]

Just successfully compile it with CUDA 10.2.
It seems that the current version just simply partition the graph by the range of rows?
also do you know is there any converter to generate the inputs for ROC like reddit-dgl?

[molamooo]

It seems that the current version just simply partition the graph by the range of rows?

Yeah I think so.

also do you know is there any converter to generate the inputs for ROC like reddit-dgl?

I read through the code of loading dataset in ROC, the format should be a simple CSC graph with indptr & indices concatenated together. It's not hard to generate a dataset by yourself, but I think the performance won't be acceptable since there is no learned-based partitioner.