Jobs on gpu022 are extremely slow

After further investigation, it looks like I was encountering the same issue as in this post, and moving everything to /scratch on the compute node as you suggested solved the issue.