A cautionary tale of configuring CPU folding slots

[bikeaddict]

Over the last few years, I've been mostly folding on GPU. I started out on CPU in the beginning before I owned any GPUs.

After acquiring some older Xeon workstations and some GPUs for folding, I switched over to various BOINC projects for CPU work. But I've occasionally tried folding on CPU again when the BOINC projects have outages or lack work to distribute.

I usually overrode the thread count setting of the CPU folding slot to use all available threads rather than leave it at -1 to configure itself. When left at -1, it excludes one thread to feed the GPU slot. But my past experience with BOINC projects and Xeons with 14 cores was that it wasn't a problem to have all the threads processing work units with no impact to the GPU thread. But the performance of the CPU folding tasks was extremely slow and some looked like they wouldn't finish before the expiration cutoff. So then I went back to BOINC for CPU tasks or just idled the CPU.

When I recently sold an older GPU and was running one system with no GPU slot and only a CPU slot, I finally had a machine completely dedicated to folding. And the processing speed was dramatically faster than previously.

So as an experiment, I configured one of the workstations running GPU tasks and added a CPU slot with thread count as -1. Now CPU tasks on that one are completing in hours rather than showing insane estimates of 2-14 days.

Now I see that folding CPU tasks cannot share a thread with other busy tasks. Lesson learned.

[toTOW]

Yes, it's how Gromacs works : it assumes that all threads are equally fast and dedicated. If one thread is slowed down by something else, the whole threads will be slowed down to the pace of the slower one ...