CPU-GPU bandwidth and latency
Currently, we have many GPU jobs that have the "wrong pinning" (or at least not the recommended one). It'd be good to know how much they lose in performance for having that. LinkTest would be an ideal program to test that, so we can compare:
- communication between a task in a given CPU core to tasks in each of the GPUs;
- communication between a task in each core with a task in a given GPU.