NaCL-support for GPU in future?

Moderators: Site Moderators, PandeGroup

NaCL-support for GPU in future?

Postby foldinghomealone » Mon Feb 13, 2017 8:46 pm

Is it planned that NaCl supports GPU folding in the future?

I think smaller WUs like CPU-NaCl-WUs which takes only a few minutes to complete are a great way to spread interest for FAH.
Even small GPU-WUs are too big and take too long for donors who are'nt as enthusiastic as others.

A GPU-NaCl would be great for many future-to-be-donors. (Smaller WUs would also help).
foldinghomealone
 
Posts: 50
Joined: Wed Feb 01, 2017 7:07 pm

Re: NaCL-support for GPU in future?

Postby bruce » Mon Feb 13, 2017 10:04 pm

FAH's GPU software depends on the presence of support for either OpenCL or CUDA.

I did a (very) little searching and wasn't able to fine a definitive statement from the people who support NaCl so it's likely the answer is currently NO (and maybe permanently so). If you locate additional information, please report it here.

Also, the general concept of creating a special family of projects that are small enough to be completed by GPUs with limited 3D calculation capabilities is a non-starter. FAH tends to deprecate low-end GPUs that have relatively few shaders.
bruce
 
Posts: 20823
Joined: Thu Nov 29, 2007 10:13 pm
Location: So. Cal.

Re: NaCL-support for GPU in future?

Postby foldinghomealone » Tue Feb 14, 2017 10:33 am

bruce wrote:FAH's GPU software depends on the presence of support for either OpenCL or CUDA.

I did a (very) little searching and wasn't able to fine a definitive statement from the people who support NaCl so it's likely the answer is currently NO (and maybe permanently so). If you locate additional information, please report it here.

Unfortunately I don't have any information about support for either OpenCL or CUDA.

bruce wrote:Also, the general concept of creating a special family of projects that are small enough to be completed by GPUs with limited 3D calculation capabilities is a non-starter. FAH tends to deprecate low-end GPUs that have relatively few shaders.

I don't think that a special family is necessary for low end GPUs. Why don't split small WUs in smaller parts and therefore have more small WUs?

Or maybe introduce a 'stop'-function that stops the WU, uploads the result, distrubute the rest-WU and give a QRB with a certain penalty.
I think this would be better than not-folding or shutting down the client and process the result at a later stage.
Last edited by foldinghomealone on Tue Feb 14, 2017 12:38 pm, edited 1 time in total.
foldinghomealone
 
Posts: 50
Joined: Wed Feb 01, 2017 7:07 pm

Re: NaCL-support for GPU in future?

Postby ChristianVirtual » Tue Feb 14, 2017 12:19 pm

My 2¥: Many smaller WU at scale would load the backend server too much from network bandwidth and processing point of view. As WU are partially sequential in nature the generation of new WU and transfer would increase with the associated cost. Would it increase the community ? Difficult to guess.
ImageImage
Please contribute your logs to http://ppd.fahmm.net
User avatar
ChristianVirtual
 
Posts: 1497
Joined: Tue May 28, 2013 12:14 pm
Location: 日本 東京

Re: NaCL-support for GPU in future?

Postby foldy » Tue Feb 14, 2017 2:37 pm

There are 3 other classes of GPUs which are too slow for the current big work units: MacOSX GPUs, Intel GPUs, Smartphone GPUs.
If there is a benefit for science in supporting slow GPUs with small work units then also NaCL webclient could be a candidate.
foldy
 
Posts: 961
Joined: Sat Dec 01, 2012 3:43 pm

Re: NaCL-support for GPU in future?

Postby rwh202 » Tue Feb 14, 2017 3:43 pm

It's also worth noting that the existing Nacl client isn't there to address slow CPUs (infact, it has rather higher CPU demands to meet the exceedingly short deadlines than the v7 client).

It is there to support more 'casual' users and deadlines are tight to minimise time wasted when users drop off before finishing WUs.

I think the intention is that the NaCl is the thin end of the wedge to entice users in with minimal effort, but the real value comes from when they are hooked and end up with 10 GTX 1080 and a $400 a month elec bill.
rwh202
 
Posts: 294
Joined: Mon Nov 15, 2010 8:51 pm
Location: South Coast, UK

Re: NaCL-support for GPU in future?

Postby foldy » Tue Feb 14, 2017 3:54 pm

10x GTX 1080 make 250 Mill PPD in a month for which you can get many FoldingCoins and CureCoins to pay the elec bill :-)
foldy
 
Posts: 961
Joined: Sat Dec 01, 2012 3:43 pm

Re: NaCL-support for GPU in future?

Postby rwh202 » Tue Feb 14, 2017 4:22 pm

foldy wrote:10x GTX 1080 make 250 Mill PPD in a month for which you can get many FoldingCoins and CureCoins to pay the elec bill :-)

Used them to buy the 10 GTX 1080s :D
The point I was trying to make was that the project appears to get the majority of it's contributions from a relatively small number of folders. The key is increasing that number of folders and to do that requires spreading the reach far and wide with the lowest barrier to entry possible, not with the expectation that the NaCl client will provide much contribution, but that it will inspire 1% on to full v7 and GPUs etc.
rwh202
 
Posts: 294
Joined: Mon Nov 15, 2010 8:51 pm
Location: South Coast, UK

Re: NaCL-support for GPU in future?

Postby ComputerGenie » Tue Feb 14, 2017 4:26 pm

rwh202 wrote:...end up with 10 GTX 1080 and a $400 a month elec bill.

I'm running 3 of them and almost 100TH/s of ASICs, a $400 electric bill would be cause for a party. :P
User avatar
ComputerGenie
 
Posts: 242
Joined: Mon Dec 12, 2016 4:06 am

Re: NaCL-support for GPU in future?

Postby foldinghomealone » Tue Feb 14, 2017 4:59 pm

Thanks for being OT all the time. Helps alot.
foldinghomealone
 
Posts: 50
Joined: Wed Feb 01, 2017 7:07 pm

Re: NaCL-support for GPU in future?

Postby bruce » Tue Feb 14, 2017 5:39 pm

I'm all for anything that enlarges the FAH commonunity providing it increases the flow of scientific results.

Breaking up a trajectory into shorter pieces an running each of those pieces wouldn't help. (A trajectory is a single group of WUs which must be completed in sequence; FAH can't assign the next Gen until they have the results of the current Gen.) If the trajectory consists of 100 pieces, each taking 24 hrs on an average GPU, the result will be available in about 3.5 months. Divide it up into 1000 pieces and assign it to a group of GPUs that are 10% as fast, and the result will be available in 3+ years -- and it's unlikely that the mobile phone will fold 24x7 (plus I'll bet the GPU, if it were supported, would be slower that the 10% number I'm using.)

FAH awards a Quick Return Bonus for a reason, and doesn't have a "No Rush" category of work.
bruce
 
Posts: 20823
Joined: Thu Nov 29, 2007 10:13 pm
Location: So. Cal.

Re: NaCL-support for GPU in future?

Postby ComputerGenie » Tue Feb 14, 2017 6:09 pm

foldinghomealone wrote:Thanks for being OT all the time. Helps alot.


My bad, I though this might be some sort of "community" type thing where people say things, like they would if they actually got to know one another on a personal level and shared a mutual comradery...
User avatar
ComputerGenie
 
Posts: 242
Joined: Mon Dec 12, 2016 4:06 am

Re: NaCL-support for GPU in future?

Postby ComputerGenie » Tue Feb 14, 2017 6:14 pm

foldinghomealone wrote:Or maybe introduce a 'stop'-function that stops the WU, uploads the result, distrubute the rest-WU and give a QRB with a certain penalty.
I think this would be better than not-folding or shutting down the client and process the result at a later stage.

Meanwhile, on topic:
So your proposal is that instead of stopping work, you should stop work? :shock:

In the larger scheme of things, I'll try to draw an analogy:
Suppose you were hitching a ride from NYC to LA...
Your suggestion is that someone should stop in Pittsburgh, pick up up, drive you 5 feet, and drop you off.
What happens if, in the time you spend getting in and out of that car, someone passes that could (and would) have driven you to straight to Barstow? Were you then helped or hindered by the 5 foot drive? And should this driver get a $20 tip for taking you that 5 feet?
User avatar
ComputerGenie
 
Posts: 242
Joined: Mon Dec 12, 2016 4:06 am

Re: NaCL-support for GPU in future?

Postby JimboPalmer » Tue Feb 14, 2017 8:12 pm

On a philosophical level, the reason there are Folding@Home donors is that the researchers are too poor to do the science themselves. So any proposal that adds costs without doing much science is a bad idea. (the android client was written by Sony, so it is 'free' to F@H)

"The Furure" is a long time. sometime in the future all GPUs will have OpenCL support and making a NACL client for them will be 'easy'. Today all Android GPUs, all Intel GPUs, and old GPUs from Nvidia and ATI/AMD fail that requirement so it is not easy.

Could a significant amount of science be done by these GPUs? Perhaps, but to be honest I bet they would provide more help to BOINC than to F@H, as BOINC is not as latency driven.
Tsar of all the Rushers
I tried to remain childlike, all I achieved was childish.
A friend to those who want no friends
JimboPalmer
 
Posts: 531
Joined: Mon Feb 16, 2009 4:12 am
Location: Greenwood MS USA

Re: NaCL-support for GPU in future?

Postby foldinghomealone » Tue Feb 14, 2017 9:15 pm

bruce wrote: Divide it up into 1000 pieces and assign it to a group of GPUs that are 10% as fast, and the result will be available in 3+ years ...

This is exactly what I'm not saying.

I reason that one 100% fast GPU + one 10% GPU is faster than a single 100% GPU.

Right now there is the situation that for slow GPUs the processing time is very long. This leads to following situations:
a) People with slow GPUs don't fold
b) Donors don't care about the QRB, because they basically don't get QRB.

Problems involved:
a) self-explanatory
b) During processing of the WU they shut down the client, continue another time or even don't continue. Therefore overall processing time is much longer compared to a situation in which the donor would upload the partly finished WU and someone else would finish it.

For sure your statement is also true. When slow GPUs tend to block and therefore lengthen the processing time of WUs then it would have negative effects like you described. Therefore you should give slow GPUs very small WUs only which they can finish in a short period similar to what you do with NaCl for CPUs.
Or let them have the possibility to upload partial results.

I think by having only quite big GPU-WUs you exclude a lot of horse power.
foldinghomealone
 
Posts: 50
Joined: Wed Feb 01, 2017 7:07 pm

Next

Return to NaCl client (Chrome);

Who is online

Users browsing this forum: No registered users and 1 guest

cron