Limiting GPU utilization percent

[jmcsnyder]

For my main question: Is there any way to limit the amount of the GPU used so an interactive window manager running through the same card retains snappy performance?

For more detail into where I am coming from:
I am new to this project and just recently got FAHClient running under Ubuntu 18.04 LTS with a primary interest in using spare GPU cycles, especially as initial packets processed suggest somewhere on the order of a 100x boost using the GPU (RTX 2080 Ti) over the CPU (Intel Xeon E5-1650v4). While the card is running up to 38C when the ambient room temp is around 20C with a 560mm radiator to cool both the motherboard and CPU through a monoblock and the GPU came with a water block (MSI Seahawk RTX 2080 Ti) and the GPU registering a 293W load on it, one of the problems is I am actively using the desktop on this computer and it gets kind of slow and jerky. This gets especially bad when only a GPU load is on it as it seems having both CPU and GPU slots active means the process controlling the GPU cannot keep the GPU as busy. Actually it seems total throughput is diminished if the CPU slot is going at the same time as the GPU slot as the GPU is so much more valuable, any GPU time lost to the controlling CPU doing other things more than negates any gains by also having the CPU do processing. It would be nice if there was a way to limit how much of the GPU gets used for Folding@Home while keeping it going. It seems the 'Light' slider setting turns off GPU processing while 'Medium' means jerky performance. It seems the only way this can work acceptably for me is if I turn off GPU processing while at the machine and hopefully remember to turn it back on when I walk away unless someone can tell me another way of handling this.

By the way in order to get the GPU working, I had to reset the GPU parameter in the 'expert' tab to true and restart the FAHClient service. I was kind of wondering why you would have this turned off by default, especially when pretty much all of the meaningful work is going to happen on GPUs?

[jonault]

I'm not aware of any setting to control the GPU utilization - AFAIK it's all or nothing. It's odd that you would be having problems with a 2080Ti. I have one also (albeit a different model without water cooling) and I don't see any slowdown or jerkiness while folding. In fact, I frequently forget to pause folding while playing World of Warcraft because I just can't tell the difference - the game plays the same either way.

[bruce]

Not really.

A powerful GPU can steal enough resources from FAH and update your screen and fold at the same time. A weak GPU cannot, so the only viable option is to configure your client to fold when the OS determines your system is idle.

That solution is not a perfect solution. Every time you wake up your computer (access the mouse or keyboard) the active WU is paused ... and then work resumes from the last checkpoint when the OS times out. It does let your computer fold all night, but if you're using the computer sporadically all day, it can be costly to FAH's throughput and it might be better to simply Pause the WU and resume when you go to bed.

[Joe_H]

By the way in order to get the GPU working, I had to reset the GPU parameter in the 'expert' tab to true and restart the FAHClient service. I was kind of wondering why you would have this turned off by default, especially when pretty much all of the meaningful work is going to happen on GPUs?

It is not actually turned off by default, but gets set that way some of the time by the installer. There are some timing issues to sort out in the next version, but basically the installer looks for a GPU and if one is not detected at that point, sets that value to false. It works more reliably on Windows, on Linux it is not as good for some reason(s).

[Paragon]

FYI, to get the GPU and CPU folding to work better together (GPU not starving), you can edit how many CPU threads you have dedicated to CPU folding. You want to keep at least one free to "feed" the GPU. That's a 6-core chip with hyperthreading, right? If your CPU slot is using all 12 threads to do CPU folding on, it will need to interrupt the work to keep feeding the 2080 Ti. Any user input on the machine further slows it down. For machines that are running folding@Home on a GPU + doing other things, I like to leave a physical core (thus two threads) free on the CPU slot.

You can manually control what is going on by editing your CPU slot and backing the # of folding threads down. To free up a physical CPU core to feed the GPU and help with backgrund tasks, go down from 12 threads to ten (don't run 11 threads because it's a large prime number and is very inefficient).

See if that helps the GPU slowdown while doing CPU folding. More info on configuring a CPU / GPU slot mixed environment is at Step 4 of this article (it's Windows, but the idea is the same)

https://greenfoldingathome.com/2020/03/ ... indows-10/

As far as having a laggy desktop, I've found this happens on some work units more than others. If you have a free PCI-E slot in the chassis, you can throw in a really low-end GPU just for the primary monitor for use when folding. This will get rid of the desktop lag.