cuGraph pain points?


Where are cuGraph users having issues and/or experiencing difficulties? For example: “doing X is slow” or “I really want to do Y, but cannot”. What are those Xs and Ys?

We have a variety of users with different goals and objectives. Their pain points vary just as widely. Some of the pain points we have observed:
• Running out of GPU memory. Graph algorithms often require auxiliary memory, and sometimes it’s not obvious a priori whether there is sufficient GPU memory
• Running in multi-node-multi-GPU can be complex. At the C++ layer we use mpirun to launch multiple processes. Running in multi-node multi GPU in python is accomplished via dask. Creating the proper environment to run these tools can be a challenge.

Access to working fast inerconnect across multi nodes like IB/NVLINK is a challenge for customers running especially in cloud environments is non trivial

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.