Welcome to MilkyWay@home

Any enhanced applications developed for Milkyway to speed up GPU work?


Advanced search

Message boards : Number crunching : Any enhanced applications developed for Milkyway to speed up GPU work?
Message board moderation

To post messages, you must log in.

1 · 2 · 3 · 4 · Next

AuthorMessage
Peter Hucker

Send message
Joined: 5 Jul 11
Posts: 199
Credit: 89,924,771
RAC: 927,984
50 million credit badge8 year member badge
Message 69439 - Posted: 18 Jan 2020, 22:43:02 UTC
Last modified: 18 Jan 2020, 23:05:23 UTC

There's apparently a SETI app (SoG?) written by a third party that runs SETI tasks much much faster provided you have Linux and a newish Nvidia card. Any such thing been done for Milkyway? Or optimizations on command line?
ID: 69439 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
alanb1951

Send message
Joined: 16 Mar 10
Posts: 47
Credit: 36,815,663
RAC: 35,699
30 million credit badge9 year member badgeextraordinary contributions badge
Message 69440 - Posted: 19 Jan 2020, 2:28:52 UTC - in response to Message 69439.  
Last modified: 19 Jan 2020, 2:54:32 UTC

There's apparently a SETI app (SoG?) written by a third party that runs SETI tasks much much faster provided you have Linux and a newish Nvidia card. Any such thing been done for Milkyway? Or optimizations on command line?

Peter,

Given that you mentioned NVIDIA cards and Linux, I suspect what you are actually thinking of is a modified version of the CUDA version of SETI@home's Multibeam application. It is impressively fast!

There is an SoG (OpenCL) application too - it stands for something like "Signals on GPU" and differs from the standard SAH OpenCL application in that it attempts to do extra processing on the GPU (and, I believe, falls back to doing that part on the CPU if it detects issues in doing so). It was also, I believe, a community enhancement like the CUDA application. Both the OpenCL applications are available for both AMD and NVIDIA, by the way...

As for enhanced applications here - I suspect the applications have been tuned as much as possible within the capabilities of OpenCL, and I don't know who might be interested in actually providing a CUDA version. (SETI@home has had CUDA versions for a long time, so there was a base-point to work from; despite that, there still isn't a viable Windows version of that program!...)

And finally, if there are command line options that can be used to hint at better use of hardware resources, I'd be interested in knowing about them too! (I guess I'm going to have to get a copy of the code to find out though - I don't recall seeing any mention here...)

A lot of GPU projects either only use OpenCL or CUDA (the latter excluding AMD users, of course) as it's a lot easier to support just one platform than both. It's a pity, but it's understandable.

Just my viewpoint, of course - I am not a MilkyWay programmer or staffer...

Cheers - Al.

[Edit.] Just been over to Einstein@home and saw your equivalent thread there. I was interested in the reply that referred to retaining expertise - definitely significant!
ID: 69440 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Peter Hucker

Send message
Joined: 5 Jul 11
Posts: 199
Credit: 89,924,771
RAC: 927,984
50 million credit badge8 year member badge
Message 69442 - Posted: 19 Jan 2020, 17:05:51 UTC - in response to Message 69440.  
Last modified: 19 Jan 2020, 17:10:08 UTC

Given that you mentioned NVIDIA cards and Linux, I suspect what you are actually thinking of is a modified version of the CUDA version of SETI@home's Multibeam application. It is impressively fast!


Correct. Never used it myself, but had conversations with people that use it.

There is an SoG (OpenCL) application too - it stands for something like "Signals on GPU" and differs from the standard SAH OpenCL application in that it attempts to do extra processing on the GPU (and, I believe, falls back to doing that part on the CPU if it detects issues in doing so). It was also, I believe, a community enhancement like the CUDA application. Both the OpenCL applications are available for both AMD and NVIDIA, by the way...


Thanks, I wasn't aware there was an OpenCL one. I only have AMD cards so that will be useful.

As for enhanced applications here - I suspect the applications have been tuned as much as possible within the capabilities of OpenCL,

And finally, if there are command line options that can be used to hint at better use of hardware resources, I'd be interested in knowing about them too! (I guess I'm going to have to get a copy of the code to find out though - I don't recall seeing any mention here...)


As far as I can tell from what others have said elsewhere, different combinations of hardware can be adjusted (in SETI anyway) with command line parameters, but because trial and error is required for every single combination, it's not done by the project and the user needs to fiddle to fine tune. I've currently got Milkyway running on my two fastest AMD cards (I got a couple of old 280X cards (from back when double precision was faster) virtually free - one works perfectly and the other needed only new fans), so that's what I'd like to speed up if possible. It's already showing me as 57th fastest in credit per day - not that I do this to boast, I want to contribute to science.

I'm not a good programmer, and it sounds like you might be, so if you can manage to find anything out about tuning MW for specific hardware, please reply here!
ID: 69442 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ProfileKeith Myers
Avatar

Send message
Joined: 24 Jan 11
Posts: 273
Credit: 130,717,335
RAC: 120,764
100 million credit badge9 year member badgeextraordinary contributions badge
Message 69443 - Posted: 21 Jan 2020, 18:36:22 UTC
Last modified: 21 Jan 2020, 18:37:24 UTC

The Nvidia CUDA app you are mentioning is euphemistically called "the special app" or "the special sauce" app. It comes in CUDA9 and CUDA10.2 flavors currently. Only usable in Linux with cards of at least C.C. 5.0 capability which means Maxwell generation onwards. It is not the SoG app. The SoG and the SAH apps are OpenCL based and can run on any card with the OpenCL API in the drivers. A couple of our GPUUG Team members are the developers. It is not released as an official stock application. Much like the older Lunatics installer optimized applications. It is somewhere between 5X to 10X faster than the stock SoG app depending on the card generation. Seti tasks are typically finished in under a minute, usually closer to 30 seconds. The application is bundled into a All-in-One BOINC package which is a ready to run BOINC platform environment once unpacked onto the Desktop or somewhere in the /home folder. Just double click the client boinc file and it is up and running. The package is at the Crunchers Anonymous website. It can be thought of as the Lunatics Installer Gen. 2.
http://www.arkayn.us/lunatics/BOINC.7z
ID: 69443 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Nick Name

Send message
Joined: 27 Jul 14
Posts: 18
Credit: 264,405,816
RAC: 1,643,211
200 million credit badge5 year member badge
Message 69445 - Posted: 22 Jan 2020, 23:08:31 UTC - in response to Message 69443.  

Nice of you folks to make this available to the public.

A CUDA app is probably not worth the effort here since Nvidia has crippled double-precision performance on their consumer GPUs. It might make sense if the double-precision requirement is dropped. There have been occasional comments about testing that but nothing has changed yet. AMD is probably going to have the edge here for a long time yet, for consumer hardware anyway.
Team USA forum | Team USA page
Always crunching / Always recruiting
ID: 69445 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Peter Hucker

Send message
Joined: 5 Jul 11
Posts: 199
Credit: 89,924,771
RAC: 927,984
50 million credit badge8 year member badge
Message 69446 - Posted: 22 Jan 2020, 23:44:47 UTC - in response to Message 69445.  

Nice of you folks to make this available to the public.

A CUDA app is probably not worth the effort here since Nvidia has crippled double-precision performance on their consumer GPUs. It might make sense if the double-precision requirement is dropped. There have been occasional comments about testing that but nothing has changed yet. AMD is probably going to have the edge here for a long time yet, for consumer hardware anyway.


Yip, I always go for double precision cards. I fail to see why they make them without it. I also detest the way Nvidia hate OpenCL. There's a wonderful language that works for all cards, yet they want their own. Reminds me of Apple.
ID: 69446 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profilemikey
Avatar

Send message
Joined: 8 May 09
Posts: 2299
Credit: 393,033,998
RAC: 35,583
300 million credit badge10 year member badgeextraordinary contributions badge
Message 69447 - Posted: 23 Jan 2020, 11:59:46 UTC - in response to Message 69446.  

Nice of you folks to make this available to the public.

A CUDA app is probably not worth the effort here since Nvidia has crippled double-precision performance on their consumer GPUs. It might make sense if the double-precision requirement is dropped. There have been occasional comments about testing that but nothing has changed yet. AMD is probably going to have the edge here for a long time yet, for consumer hardware anyway.


Yip, I always go for double precision cards. I fail to see why they make them without it. I also detest the way Nvidia hate OpenCL. There's a wonderful language that works for all cards, yet they want their own. Reminds me of Apple.


Because we crunchers are such a small part of their business they don't care if buy them or not. Their business is super computers and gamers, neither needs dual precision so it's still there but a smaller and smaller part of it so they can give gamers and super computers what they want so they will buy the latest and greatest gpu's. Putting more emphasis on dual precision raises the price of the cards that they don't see a return on.
ID: 69447 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Peter Hucker

Send message
Joined: 5 Jul 11
Posts: 199
Credit: 89,924,771
RAC: 927,984
50 million credit badge8 year member badge
Message 69448 - Posted: 23 Jan 2020, 14:28:12 UTC - in response to Message 69447.  

Because we crunchers are such a small part of their business they don't care if buy them or not. Their business is super computers and gamers, neither needs dual precision so it's still there but a smaller and smaller part of it so they can give gamers and super computers what they want so they will buy the latest and greatest gpu's. Putting more emphasis on dual precision raises the price of the cards that they don't see a return on.


There are some cards available (more expensive ones) with better dual precision. I assume those must be aimed at us and/or scientists.

I wonder if games use double precision at all?

I take it bitcoin isn't double? That's a pretty big market.
ID: 69448 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
alanb1951

Send message
Joined: 16 Mar 10
Posts: 47
Credit: 36,815,663
RAC: 35,699
30 million credit badge9 year member badgeextraordinary contributions badge
Message 69449 - Posted: 24 Jan 2020, 3:10:03 UTC - in response to Message 69448.  

@Peter Hucker
I take it bitcoin isn't double? That's a pretty big market.

Bitcoin operations use integer arithmetic - it's all about hashing functions... (And that's why AMD cards were more popular than NVIDIA for Bitcoin, by the way...)

Cheers - Al.
ID: 69449 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ProfileKeith Myers
Avatar

Send message
Joined: 24 Jan 11
Posts: 273
Credit: 130,717,335
RAC: 120,764
100 million credit badge9 year member badgeextraordinary contributions badge
Message 69450 - Posted: 24 Jan 2020, 7:32:41 UTC - in response to Message 69446.  

Yip, I always go for double precision cards. I fail to see why they make them without it. I also detest the way Nvidia hate OpenCL

They do make double-precision optimized cards, they are called Quadros and Teslas. They charge a pretty penny for them too since the target clientele are high-speed Wall Street traders and computational scientists.
OpenCL is a generalized computing API. Not best for any one thing, and just mediocre at all things. So why further develop a mediocre programming API when you developed the first "race-horse" computing API called CUDA. The performance of CUDA over OpenCL is not even close. They have plenty of customers who are willing to pay for a thoroughbred. So their focus is to continue to develop CUDA. They joined the OpenCL consortium out of the desire to appear to not block OpenCL and not appear to try and force CUDA down any one else's throat. So there is an alternative choice if so desired.

The better double-precision math of AMD cards compared to Nvidia for the consumer cards has nothing to do with the programming API but rather in the choices of how to configure the silicon of their gpus for their target gamer audience.
ID: 69450 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Peter Hucker

Send message
Joined: 5 Jul 11
Posts: 199
Credit: 89,924,771
RAC: 927,984
50 million credit badge8 year member badge
Message 69452 - Posted: 24 Jan 2020, 19:57:07 UTC - in response to Message 69450.  
Last modified: 24 Jan 2020, 20:01:51 UTC

Yip, I always go for double precision cards. I fail to see why they make them without it. I also detest the way Nvidia hate OpenCL

They do make double-precision optimized cards, they are called Quadros and Teslas. They charge a pretty penny for them too since the target clientele are high-speed Wall Street traders and computational scientists.
OpenCL is a generalized computing API. Not best for any one thing, and just mediocre at all things. So why further develop a mediocre programming API when you developed the first "race-horse" computing API called CUDA. The performance of CUDA over OpenCL is not even close. They have plenty of customers who are willing to pay for a thoroughbred. So their focus is to continue to develop CUDA. They joined the OpenCL consortium out of the desire to appear to not block OpenCL and not appear to try and force CUDA down any one else's throat. So there is an alternative choice if so desired.

The better double-precision math of AMD cards compared to Nvidia for the consumer cards has nothing to do with the programming API but rather in the choices of how to configure the silicon of their gpus for their target gamer audience.


How do I tell which projects use CUDA? I assume the Nvidia icon here could mean using OpenCL: https://boinc.berkeley.edu/projects.php

Mind you, Nvidia cards are prohibitively expensive, so I think I'll stick with AMD.
ID: 69452 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ProfileKeith Myers
Avatar

Send message
Joined: 24 Jan 11
Posts: 273
Credit: 130,717,335
RAC: 120,764
100 million credit badge9 year member badgeextraordinary contributions badge
Message 69453 - Posted: 24 Jan 2020, 21:28:21 UTC - in response to Message 69452.  
Last modified: 24 Jan 2020, 21:31:51 UTC

How do I tell which projects use CUDA? I assume the Nvidia icon here could mean using OpenCL: https://boinc.berkeley.edu/projects.php

Actually, the Nvidia logo on the project lists denotes gpu applications, NOT necessarily OpenCL applications nor CUDA applications. You would still have to look at the https://{project]/apps.php page to see if their are CUDA applications.
ID: 69453 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Peter Hucker

Send message
Joined: 5 Jul 11
Posts: 199
Credit: 89,924,771
RAC: 927,984
50 million credit badge8 year member badge
Message 69456 - Posted: 24 Jan 2020, 21:43:02 UTC - in response to Message 69453.  

How do I tell which projects use CUDA? I assume the Nvidia icon here could mean using OpenCL: https://boinc.berkeley.edu/projects.php

Actually, the Nvidia logo on the project lists denotes gpu applications, NOT necessarily OpenCL applications nor CUDA applications.


As I thought, which is what I was suggesting, that the info isn't on my linked page.

You would still have to look at the https://{project]/apps.php page to see if their are CUDA applications.


Seems Milkyway and Einstein are both OpenCL only. I guess programming for CUDA aswell is too much hassle?
ID: 69456 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ProfileJoseph Stateson
Avatar

Send message
Joined: 18 Nov 08
Posts: 215
Credit: 1,153,865,024
RAC: 35,110
1 billion credit badge10 year member badge
Message 69458 - Posted: 25 Jan 2020, 5:04:04 UTC - in response to Message 69456.  
Last modified: 25 Jan 2020, 5:05:49 UTC


Seems Milkyway and Einstein are both OpenCL only. I guess programming for CUDA aswell is too much hassle?


Pretty sure that OpenCL has all sources available.
OTOH, NVidia only made the API available.

OpenCL is open source.
CUDA is open standards.

However, I suspect the real difference is the OpenCL app can run on amd, intel and nvidia systems, unlike CUDA: limited to nvidia boards.
Perhaps even ARM could support OpenCL.
ID: 69458 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Peter Hucker

Send message
Joined: 5 Jul 11
Posts: 199
Credit: 89,924,771
RAC: 927,984
50 million credit badge8 year member badge
Message 69459 - Posted: 25 Jan 2020, 12:07:23 UTC - in response to Message 69458.  
Last modified: 25 Jan 2020, 12:09:01 UTC


Seems Milkyway and Einstein are both OpenCL only. I guess programming for CUDA aswell is too much hassle?


Pretty sure that OpenCL has all sources available.
OTOH, NVidia only made the API available.

OpenCL is open source.
CUDA is open standards.

However, I suspect the real difference is the OpenCL app can run on amd, intel and nvidia systems, unlike CUDA: limited to nvidia boards.
Perhaps even ARM could support OpenCL.


I think they're using something called "pie" - these are the four projects which run on my ARM phones (the only four projects which ever provide ARM work - and only Universe and Yoyo provide a consistant amount):

https://lhcathome.cern.ch/lhcathome/apps.php
http://www.asteroidsathome.net/boinc/apps.php
https://www.rechenkraft.net/yoyo/apps.php
https://universeathome.pl/universe/apps.php
ID: 69459 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ProfileKeith Myers
Avatar

Send message
Joined: 24 Jan 11
Posts: 273
Credit: 130,717,335
RAC: 120,764
100 million credit badge9 year member badgeextraordinary contributions badge
Message 69461 - Posted: 26 Jan 2020, 3:16:14 UTC - in response to Message 69459.  

Actually, the very first CUDA application for Seti was written by actual Nvidia developers and programmers.

It is up to the individual project scientists to develop CUDA applications or have the project volunteers do it for the project. Since most science projects operate on a shoe-string budget, there are no resources to fund a project programmer.

The fact that we have two volunteer GPUUG team members competent in CUDA programming allowed us to take the legacy CUDA code developed by Nvidia and early Seti project volunteers and morph it into the present special app CUDA10 application. That is why the CUDA10 application can use all of the present day hardware capabilities and is so impressively fast.
ID: 69461 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Peter Hucker

Send message
Joined: 5 Jul 11
Posts: 199
Credit: 89,924,771
RAC: 927,984
50 million credit badge8 year member badge
Message 69462 - Posted: 26 Jan 2020, 12:32:48 UTC - in response to Message 69461.  

Actually, the very first CUDA application for Seti was written by actual Nvidia developers and programmers.

It is up to the individual project scientists to develop CUDA applications or have the project volunteers do it for the project. Since most science projects operate on a shoe-string budget, there are no resources to fund a project programmer.

The fact that we have two volunteer GPUUG team members competent in CUDA programming allowed us to take the legacy CUDA code developed by Nvidia and early Seti project volunteers and morph it into the present special app CUDA10 application. That is why the CUDA10 application can use all of the present day hardware capabilities and is so impressively fast.


Yes I guess CUDA is probably designed more for people running big things like Adobe Photoshop, Premiere, etc. Adobe has money for programmers.

But I still think Nvidia has made a bad choice. Compatibility is key - as you've said shoe-string budget places ain't gonna write the code twice. I have always and will always buy AMD. The first thing that put me off was a brand new Nvidia card 5 years ago which melted after 30 minutes use! It had no thermal cutout or throttling!
ID: 69462 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ProfileJoseph Stateson
Avatar

Send message
Joined: 18 Nov 08
Posts: 215
Credit: 1,153,865,024
RAC: 35,110
1 billion credit badge10 year member badge
Message 69463 - Posted: 26 Jan 2020, 13:39:19 UTC - in response to Message 69462.  

The first thing that put me off was a brand new Nvidia card 5 years ago which melted after 30 minutes use! It had no thermal cutout or throttling!


You got this one right, AMD has been known for lower power requirements for same type of competing card as long as I can remember.

When NVidia comes out with a new card they give it a new name gtx670 => 770 but it is just the same 104 chip with higher speed and they enabled more features which, of course, runs a lot hotter.

What came after the gtx-460 ? They simply added an extra chip and two more fans as the original 460 was too hot even for one fan.

Apple stopped using Motorola CPUs when Intel showed they easily could meet power requirements.

With NVidia, it is marketing. However, their investment in CUDA has paid off for the scientific community and OpenCL was a late attempt to join the bandwagon but it did provide competition (diversity) which helps bring down prices.

I use the following CUDA app regularly at home
seti & gpugrid of course
DVDfab for ripping. You would not want to rip a bluray without CUDA.
Splashtop for remote desktop really slow on AMD boards
Very rarely that Adobe you mentioned, or Nero as I don't to much video encoding anymore.
as far as l know they do not do OpenCL but I do have both products.

However, the research company I retired from used CUDA when it first came out for all types of applications and ended up switching to Linux and away from Windows. CUDA was a contributing factor.
ID: 69463 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Peter Hucker

Send message
Joined: 5 Jul 11
Posts: 199
Credit: 89,924,771
RAC: 927,984
50 million credit badge8 year member badge
Message 69464 - Posted: 26 Jan 2020, 14:11:45 UTC - in response to Message 69463.  
Last modified: 26 Jan 2020, 14:12:35 UTC

You got this one right, AMD has been known for lower power requirements for same type of competing card as long as I can remember.


Oh I don't know. The two AMD R9 280X I'm using for MW just now - one has 2 fans, one has 3, and they run at 70C with full speed fans all the time in a normal full tower case. I have since attached them further from the motherboard using risers to give them more airflow, left the side off the case, and added plenty of 120mm and 140mm fans on the case to blast cold air across them. 250W TDP is way too much heat to get rid of from one chip. They now run at 60C with 60% fan.

But.... if you overheat an AMD card, it throttles or shuts off. The brand new Nvidia didn't, it just broke. I sent it back and insisted they replace it with an equally priced AMD.

What came after the gtx-460 ? They simply added an extra chip and two more fans as the original 460 was too hot even for one fan.


The AMD 7990 I think is simply two 7970s on one card. I don't know if anybody's ever managed it, but I never got it to run at 100% without overheating. That was it's own 3 fans at 100%, plus a 16 inch fan from a Range Rover 3.5L V8 car radiator(!) blowing directly over it. Still couldn't use 100% clock for bitcoin mining without it getting worryingly hot (as in approaching 90C in a 20C room).

I use the following CUDA app regularly at home
seti & gpugrid of course
DVDfab for ripping. You would not want to rip a bluray without CUDA.


Depends how many you do, I've ripped them without a GPU at all. I just leave it doing it overnight or while I do something else. Of course if you want 10 ripped in a day....

Splashtop for remote desktop really slow on AMD boards


That's just a remote desktop right? Works fine for me, I use "Remote Utilities". The built in Windows 10 one seems to switch off GPU computing for some reason. I wasn't aware you needed GPU computing for something that simple.
ID: 69464 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ProfileKeith Myers
Avatar

Send message
Joined: 24 Jan 11
Posts: 273
Credit: 130,717,335
RAC: 120,764
100 million credit badge9 year member badgeextraordinary contributions badge
Message 69465 - Posted: 27 Jan 2020, 1:58:45 UTC - in response to Message 69462.  

I have always and will always buy AMD.

That is certainly your prerogative as a consumer.

In my case, I will never buy an AMD gpu because the drivers have been, are, and always will be a nightmare to install.

Nvidia on the other hand . . . download, install and boom, up and running in minutes. And never have to struggle with them ever again.
ID: 69465 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
1 · 2 · 3 · 4 · Next

Message boards : Number crunching : Any enhanced applications developed for Milkyway to speed up GPU work?

©2020 Astroinformatics Group