GPU Memory usage low (1%) volatility high (98%)

[Cathal]

Hi there, I'm new to F@H.

I have a 1080 Ti and a 12 core i7 rig. Currently I'm running project 11761 - Coronavirus SARS-CoV-2. However it is using less than 200mb of an available 12GB, and is using that at a very high utility. This seems very ineffecient to me and I was wondering if this was normal, or if there was something problematic with my setup.

Thanks in advance.

[bruce]

In efficient in what way? An application that needs no more than 200Mb need not allocate more, contrary to what the RAM vendors would have you believe. On the other hand, an application that can use all available CPU resources for computing should be doing that. FAH will use 100% of that resource at a very low priority so that if you run foreground activites (say a browser, it will get the necessary processing resources first (at a higher priority) so it shouldn't interfere with your use of the computer.

If running at 100% uses more power/generates more heat than you'd like or runs your battery down, you can reduce that activity by adjusting the number of CPU cores allocated to FAH

[JimboPalmer]

This is a CPU project
https://apps.foldingathome.org/project?p=11761
He/she is looking at a protease, a small protein. (A large Protein would be better served by a GPU project but it is never worth while to throw more Processing Units at Folding than you have atoms)
https://en.wikibooks.org/wiki/Structura ... /Proteases

So at most they might need 3 copies of the atom database they are working on. (From, Now, and To) This not much memory compared to larger Proteins. But as the program simulates the atoms folding around each other through many time slices, that memory will change a lot.

So, if it needed more memory, it would be because it had more atoms, and would become a GPU job. It is a CPU project because it is doing an in depth study of a smaller molecule. F@H itself rarely uses a gig of memory because if it did it would be a GPU project. Even GPUs rarely have memory issues, gamers need more texture memory than folders need.

[bruce]

[quote="JimboPalmer"]... that memory will change a lot../quote]

Great explanation, pseudo-7im

First time i read it I interpreted those words as ... that memory allocation would change a lot... (which is false) and then I realized you meant
... the values in those memory locations would change a lot...

Sometimes it pays to add another word or two to the explanation just to be sure some idiot reading it understands what you mean.

[Cathal]

In response to bruce I'm talking video ram, not my actual ram. I'm also not providing CPU compute, but GPU compute so I'm not sure if your response still applies.
Also as an aside, medium and high setting make no difference in memory and volatility usage, it uses the same amount of gpu memory at a very high utility, which I'm not super comfortable with.

to JimboPalmer, I'm coming at this from a Deep Learning batch wise POV, rather than a gamer if that helps.

I guess my first question is that if I dedicate only GPU resources, why am I being given a CPU project?

The second question I have is that if F@H rarely uses more than a couple hundred megs per protein, why not have multiple proteins on the same GPU at the same time, much the same way that multiple deep learning graphs can exist on the GPU?

[bruce]

Each GPU requires the resources of one CPU thread to move data to/from the GPU. The other CPU resource are avilable and normally get allocated to run a different Work Unit

The OpenMM/OpenCl code does a pretty good job of keeping a large percentage of the shaders busy computing floating point instructions in parallel. Other GPU resources remain mostly idle. Running multiple proteins concurrently is not supported. It would add to the complexity and (probably) slow down each of the analysis Kernels. If you have the developer toolkit you can probably examine what's happening in your GPU and tell me if my assumptions apply in your case.

[Cathal]

To be clear I didn't limit any CPU access, I just selected the gpu radio button.

Regarding my question about medium vs high usage, is there a reason there appears to be no difference in performance?

I'm viewing gpu performance via Nvidia-smi.

[JimboPalmer]

Again, F@H is almost never memory size constrained or memory speed constrained, it is almost always constrained by Processing power. They do not switch to GPU folding to get more memory or memory speed, they get more Processors!
A top end Graphic card may have 4000 'CUDA Cores' while the best micro processors have 256 threads. (and 16 threads is a lot more common)

(as an aside, Cache size does help F@H as the 'interesting' part of the program can stay in cache and only rarely go to main memory)

"Regarding my question about medium vs high usage, is there a reason there appears to be no difference in performance? "

High uses all the CPUs, and Medium uses all but one.

This is an artifact of CPU designs improving faster than the program changes.

When we had only 4 Cores, using 4-1 meant we only used 3/4 of the CPUs, with 12 cores 12-1 means that we still use 11/12ths of the CPUs, it is a minor difference from High.

[Cathal]

Hmm ok. I think I follow.

One thing that's still bothering me is that the volatility is a bit too high for my comfort and it seems to be causing a slowdown/blackouts. It could be the joys Nvidia and Ubuntu not playing nice, I'm not sure.

I think I'll sooner switch to CPU as I don't seem to have much control over gpu usage. Which is a pity.

[Porkey]

I think there's been a misunderstanding in this thread.

I'm not sure what you mean by GPU "volatility." After looking at a couple Google results on nvidia-smi, I think you may be mistaking a "volatility" reading for a "GPU usage" or "GPU utilization" reading?

I've only been a sort-of casual F@H participant since 2013, but in all my years of folding, my understanding of your situation is completely normal. F@H work units do not consume any noticeable amount of VRAM by today's standards.
What F@H is looking for in a GPU is the GPU itself. As Jimbo explained, GPUs technically have thousands of cores, and at the end of the day are usually much better suited for F@H's kind of work because they are faster at this kind of stuff.
F@H needs the processing power; not copious amounts of RAM.

Therefore it is normal to see very low V/RAM usage, and very high (near 100%) GPU utilization as the GPU core churns away at the work.

This is a CPU project
https://apps.foldingathome.org/project?p=11761

p11761 is a Core22 project, which means it is a GPU project, doesn't it? I don't follow FAHCore development too closely, but last time I checked, Core22 was solely for doing work on a G-PU, and not a C-PU?
I'd already noticed this, but, I think, in all their haste to pump out as many work units as possible... Someone might just be copy/pasting project descriptions..?

[JimboPalmer]

Core_22 is for GPUs.

Perhaps Neha Vithani was tired and confused when the description was written.

I have no idea why it worries Cathal that the memory contents are changing rapidly.