nvGraph multi-gpu support


I am trying to use nvGraph library on mult-gpu environment. My use case requires to perform triangle counting on a very large graph that cannot fit in single GPU memory. Can nvgraph library triangle counting api default support multi-gpu here or should I partition the graphs? If I should partition the graph, then I should i perform the reduction?

Any thoughts?

nvgraph does not currently support the use of multiple GPUs per API call.

Can I still partition it and use it? (this is possible only if I know what is the algorithm it uses. However, I dont see any documentation on it. So, I am not sure!)

I believe since the algorithm(s) is unpublished, the safe assumption is that it could change. For correctness, then, you should manually count the triangles broken by the partition(s).

It’s not clear to me why counting the number of triangles is algorithm-dependent. The number of triangles is a property of the graph; presumably there is only one correct answer.

Presumably, to answer your last question, you should perform the reduction yourself.