Bigger Job Than my GPU Can Handle

[Starduster]

Hi. I had assumed that the Folding @ Home software could accurately determine the capabilities of the machine running its software and assign jobs that the computer could finish before expiration. For most of the year & a half I have been running F@H, that's been the case. Yet recently, F@H has been assigning some jobs that my AMD Radeon Pro WX-2100 graphics card cannot possibly finish before deadline, and 1 &1/2 day's effort gets dumped at 70-80% completion, no point credit. As of right now, my GPU is running Project 18450, deadline in 1.64 days, my estimated completion is 2.08 days. The issue has never happened with a CPU job.

I am running current Radeon Pro drivers tweaked to perfection, and I am generating a 50,000 PPD average off the GPU. Since the computer is my daily driver, setting the AMD software to optimize for memory over graphics is a no go. Compromises my desktop and its functionality. I have it running as good as it can.

Has anyone else suffered this issue, and is there a workaround at the F@H end?

Merci.

[JimboPalmer]

Welcome to Folding@Home!

There is an algorithm to match Work Units to your GPU, but it is based on it's capability, not it's speed. So a GT 1030 gets the same WUs as a GTX 1080 as they only differ in speed, while a GTX 980 might get a different WU, as they are different generations.

I am on vacation and am not near my notes, (hopefully someone will correct my advice) but you can set MAX-PACKET-SIZE to SMALL and get smaller files, which tend to be shorter run times. (This was designed fo dial up modems, so is rarely used)

[BobWilliams757]

Without looking at the GPUs.txt file, my guess is that your GPU was put into a species that is too high for the capabilities. This often happened on my Vega 11 onboard graphics. The only legit workaround is to note the projects that your system won't make a deadline on, then look up the cause. If you are lucky (it happens) and the offending work units are all related to one cause, then simply select another cause under your settings. This can help avoid the problem projects, but isn't a guarantee.

Other than that, you can try to wait it out and CPU fold, but depending on the project that can take a lot of time. F@H does allow you to dump a work unit if your system can't complete it in time, so at least you aren't stuck waiting on something to finish unless you want. I tried letting them run until another user had completed them after the deadline but before expiration, but in most cases they got picked up and finished by another folder.


I will say that overall the species assignments are lacking on the AMD gear more than the Nvidia. I understand this is somewhat a product of their structures and such, and it's harder to do. But i think we could do much better, and I'd gladly contribute some time into helping fix it if those "in the know" can point me in the right direction. The waste of resources is greater if we don't fix it, and some of those older slower cards still run quite well on small atom count projects.

[JimboPalmer]

https://www.techpowerup.com/gpu-specs/r ... 2100.c2979

[tvdsluis]

Has anyone else suffered this issue, and is there a workaround at the F@H end?

Yes i have.
It happens on both AMD and NVIDIA.
It's a shame because everything after the timeout is a waste of energy and money.
They should set the timeout at 2 days with the current algorithm to work correctly.
Cards that can finish some projects easily on time, get assigned projects they can't finish on time.
Weird and a waste.

[tvdsluis]

I am on vacation and am not near my notes, (hopefully someone will correct my advice) but you can set MAX-PACKET-SIZE to SMALL and get smaller files, which tend to be shorter run times. (This was designed fo dial up modems, so is rarely used)

I have tried that on different machines, and unfortunately that doesn't work.
I have seen 2 mb packets running 48 hours well past the timeout, and 65mb packets running 12 hours.
I have changed this setting a year ago, so i am pretty confident it doesn't help.

[Starduster]

Quick addendum- The size of the assigned job isn't the issue, it is the points per day pace that job is asking of the GPU in order to complete it on time.

[BobWilliams757]

Quick addendum- The size of the assigned job isn't the issue, it is the points per day pace that job is asking of the GPU in order to complete it on time.

It's the unfortunate reality of how GPU's are classified and the assignment process. I found myself in a simlar place. I didn't care if a work unit took days, as long as it helped. But they started getting assignments that it wouldn't complete on time, and that is frustrating. The shame is that with some assignment specifics worked out, a lot of the cards with less shaders still run great on certain work units.