Could smp folding support GPGPU?

[pipnina]

for example. my i5 2500k gets between 3000 and 20,000 ppd depending on size of wu. i dont like to use GPU folding because it make watching you tube videos impossible.
i was wandering if my gtx 560ti could use its cuda technology to "boost" the cpu's folding performance by performing some of the calculations...

[rhavern]

Not yet, but perhaps in the future, http://folding.typepad.com/news/2012/09 ... w-you.html

[bruce]

Welcome to foldingforum.org, pipnina.

The FahCore_15 and FahCore_16 are explicitly written using GPGPU methodology so your title is not an accurate description of what I think you're actually looking for. There are a couple of unsolved problems involved in a complete answer to your questions.

First, consider the concept of "some of the calculations." The software would have to decide which parts of the calculation to send to the CPU and which parts to send to the GPU. That sort of decision would depend a lot on your hardware. Does one person have a fast CPU and a relatively slow GPU and another person have a slower CPU and a faster GPU? The same division of work between two devices wouldn't work well for both people. Finding the right division for everybody is a very, very complex task.

Second, current GPUs and their supporting software has no concept of priority. GPGPU tasks to fold a protein and to decode a youtube video are processed in first-come-first-served order. If the GPU is (mostly) doing one thing, that's not unreasonable, but if you want folding to wait whenever there's video work to be done, somebody has to figure out how to make that happen. I have no doubt that something like this will be available in a few years, but I'm pretty confident that it won't happen with the current hardware and software. GPGPU still needs a lot of development in this area.

Third, you're asking for the division of labor to be dynamic. Your GPU would be "fast" most of the time, but when you're watching youtube videos, it's busy doing something else and without those resources sending a fixed amount of work to the GPU would find your CPU with free time on it's hands while the GPU work waits for the GPU to process its share of the work. The current version of GROMACS does contain some software that can do dynamic load balancing but (apparently) it is still much more primitive that would be required. So far I've only seen it used to balance blocks of calculations which are assigned to CPU cores which are symmetric (i.e.-the hardware is well balanced by being identical although the blocks of work may not be precisely matched). That's what the "S" in "SMP" means.

Dr. Pande's recent blog post (referenced above) seems to imply that development work has been going on to improve the current limitations. I'm sure he knows more about the current state of those efforts than I do, but I do want to emphasize that it's a very complex problem and one that's on the forefront of technology. When it does reache the "it works" stage of development, I expect that you'll see it first in FAH.

[P5-133XL]

I'm not sure what difference it would make to YouTube videos. The problem with the GPU making videos stutter is that you have multiple applications using the GPU. If you appended GPU folding to SMP folding you'd still be in the same situation in regards to the tube. If you can't stand it now, what makes you think you'd tolerate it then?

My suggestion is do like I do, use the pause on your GPU slot when some non-folding application needs to use the GPU and fold the rest of the time on the GPU.

[pipnina]

If you appended GPU folding to SMP folding you'd still be in the same situation in regards to the tube. If you can't stand it now, what makes you think you'd tolerate it then?

the problem is. i cant pause them independently. as i use the client control. and SMP folding does not cause stuttering on videos, or result in slowdown in games to my knowledge. and i don't like the older versions of the client, as they do not specify ppd estimates. or give an explanation of the clients running. this is interesting, as through the new client, i have learned my i5 2500k likes medium sized proteins, too large and the point bonus does not compensate for the time taken to process, too small and the simulation takes longer than it could. the base points is about 800 for my cpu to make the most of it.
my gpu however, because i cant use it for folding without sacrificing real performance. plus the lack of gpu point bonuses.

if folding used cuda for example. (which i believe is nvidia's gpgpu solution) the processes that take longer on cpu's get dumped on the gpu? i think thats how gpgpu works?

[Zagen30]

the problem is. i cant pause them independently. as i use the client control.

Right-clicking a slot in the Folding Slots pane of FAHControl allows you to control slots individually (run/pause/finish). I admit it's not very obvious, especially when there's the big buttons up top grabbing one's attention.

[Joe_H]

You do need to be in Advanced or Expert mode for the Folding Slots pane to show in FAHControl. Then you can right-click on a slot and just pause that instead of all slots like the big Pause button on top.

[Jesse_V]

the problem is. i cant pause them independently. as i use the client control.

Right-clicking a slot in the Folding Slots pane of FAHControl allows you to control slots individually (run/pause/finish). I admit it's not very obvious, especially when there's the big buttons up top grabbing one's attention.

V7 has to be in Advanced mode to give you these options. The mode can be changed using the dropdown menu in the upper righthand corner.

[pipnina]

This is something i didn't know before. thank you for sharing.
however. if the SMP folding should have the same effect as the GPU folding in terms of flash applications / games / watching videos. then why doesn't it for me? my is is running on SMP:4 because hyper threading is disabled on core i5's.
i am not overclocking, in fact, the opposite, because for some unknown reason my processor will not turbo past 3.4GHz when folding. i have a decent cooler (corsair h80) so it shouldn't be throttling.
help is much appreciated.

[7im]

the problem is. i cant pause them independently. as i use the client control.

Right-clicking a slot in the Folding Slots pane of FAHControl allows you to control slots individually (run/pause/finish). I admit it's not very obvious, especially when there's the big buttons up top grabbing one's attention.

V7 has to be in Advanced mode to give you these options. The mode can be changed using the dropdown menu in the upper righthand corner.

Actually, it can be in either Advanced mode or Expert mode. And I think they call that specific format a drop down list, not a drop down menu.


For me, SMP folding does not affect video like GPU folding does.

Also note that folding uses a processor's resources very heavily, which creates more heat. So it may not turbo as high. What is the top turbo speed for that chip?

[bruce]

This is something i didn't know before. thank you for sharing.
however. if the SMP folding should have the same effect as the GPU folding in terms of flash applications / games / watching videos. then why doesn't it for me? my is is running on SMP:4 because hyper threading is disabled on core i5's.
i am not overclocking, in fact, the opposite, because for some unknown reason my processor will not turbo past 3.4GHz when folding. i have a decent cooler (corsair h80) so it shouldn't be throttling.
help is much appreciated.

SMP (or uniprocessor) projects will never cause the same sort of interference as GPGPU processing. Your Operating System (Windows, in this case) has years of development and includes some pretty sophisticated internal code that assigns priority to each task. Depending on both the magnitude of the request for resources and the priority of the task making the request, Windows allows the programmers to establish a very low priority for FAH so that whenever something else needs to be done, it gets immediate access to the processor.

It will be many years before GPGPU develops the sophistication to do that. Instead, it processes everything on a first-come-first served basis. When I move my mouse, if the GPU is idle and if Windows chooses to use the GPU to update the mouse pointer on my screen, it happens instantaneously. If FAH has just submitted a block of GPU work that takes one second for the GPU to process, I won't see that the mouse has moved until one second passes. If less than one second of video has been prepared in a form that it can be displayed by just using the CPU, I'll see video stuttering until the GPU can decode a longer piece of the video stream.

There's some room to adjust the "one second" Making that a larger number makes FAH-GPU more efficient (within reason) and making a smaller number improves the ability of the GPU to keep ahead of the flash player but the fundamental fact is that the flash player cannot be assigned a higher priority than GPGPU work, nor can either one interrupt the other before it finishes whatever it's been assigned. Programmers for flash, for games, for Windows, for FAH are all faced with this problem and for the most part, they simply assume that the GPU will almost never be busy doing something else. That assumption is false when FAH-GPU is running. (Programmers who write code for the CPU are not faced with this problem.)

[pipnina]

Also note that folding uses a processor's resources very heavily, which creates more heat. So it may not turbo as high. What is the top turbo speed for that chip?

my i5 2500k should turbo to 3.7 GHz and it has a stock speed of 3.3 GHz, when folding, it only turbos to 3.4 (ish).

i have a corsair h80 closed loop water cooler, and at maximum load (like now) when running at 3.4 GHz, i only get temps of 50 degrees Celsius on the hottest core (core 0, which is usually the hottest) and when i had my last 2500k overclocked to
5 GHz, i only got temps going up to 75c
so i don't see why it wont turbo to 3.7, however, when it does in other programs, i have tried to change things in my new motherboard, but the mouse wont work properly in the gigabyte 3D bios.

[P5-133XL]

Turbo clock rates are determined by the number of cores in use, not temperature. While folding, all cores are in operation and as such the turbo speed is significantly lower than if you are running an application that is only using a single core.

[pipnina]

Turbo clock rates are determined by the number of cores in use, not temperature. While folding, all cores are in operation and as such the turbo speed is significantly lower than if you are running an application that is only using a single core.

but. that does not make sense, if all cores are at 100% load, wouldn't that be a reason to increase the turbo frequency? and considering i have a good 20 degrees before the cpu starts getting uncomfortably hot, it should be at 3.7 at the moment.

[artoar_11]

my i5 2500k should turbo to 3.7 GHz and it has a stock speed of 3.3 GHz, when folding, it only turbos to 3.4 (ish).

i have a corsair h80 closed loop water cooler, and at maximum load (like now) when running at 3.4 GHz, i only get temps of 50 degrees Celsius on the hottest core (core 0, which is usually the hottest) and when i had my last 2500k overclocked to
5 GHz, i only got temps going up to 75c
so i don't see why it wont turbo to 3.7, however, when it does in other programs, i have tried to change things in my new motherboard, but the mouse wont work properly in the gigabyte 3D bios.

If lag is vexatious, stop GPU slot temporarily.
5 GHz is high CPU frequency for FAH. You can work very well with frequencies between 4.4 - 4.7 GHz.