SETI@home v8.22 Windows GPU applications support thread

Message boards : Number crunching : SETI@home v8.22 Windows GPU applications support thread
Message board moderation

To post messages, you must log in.

Previous · 1 · 2 · 3 · 4 · 5 · 6 . . . 7 · Next

AuthorMessage
Kets_One

Send message
Joined: 5 Jun 03
Posts: 9
Credit: 972,549
RAC: 2
Netherlands
Message 1859604 - Posted: 5 Apr 2017, 15:21:49 UTC
Last modified: 5 Apr 2017, 15:23:36 UTC

Guys,

Ive managed to find useable info on app_config.xml and created the file in SETI directory.
Ive tested some settings and its now working as intended.

However, i have no clue on further arguments in app_config.xml, or in mb_cmdline_win_x86_SSE3_OpenCL_NV_SoG.txt or in mb_cmdline_win_x64_AVX_VS2010.txt....

BTW: i currently get around 600 credits/day for the work computer...

Thanks,
Kets
ID: 1859604 · Report as offensive     Reply Quote
Grant (SSSF)
Volunteer tester

Send message
Joined: 19 Aug 99
Posts: 8896
Credit: 115,387,733
RAC: 70,393
Australia
Message 1859654 - Posted: 5 Apr 2017, 18:30:34 UTC - in response to Message 1859586.  
Last modified: 5 Apr 2017, 18:32:40 UTC

Intel i7-4500U (SSE, AVX) (i dont use the CPU for SETI)
Nvidia GT 750M (i use 100% for SETI)
16GB system ram

Using all the CPU cores with the AVX application would outperform that GPU.
Grant
Darwin NT
ID: 1859654 · Report as offensive     Reply Quote
Stephen "Heretic"Project Donor
Volunteer tester
Avatar

Send message
Joined: 20 Sep 12
Posts: 2638
Credit: 48,865,133
RAC: 138,286
Australia
Message 1859682 - Posted: 5 Apr 2017, 20:47:47 UTC - in response to Message 1859654.  

Intel i7-4500U (SSE, AVX) (i dont use the CPU for SETI)
Nvidia GT 750M (i use 100% for SETI)
16GB system ram

Using all the CPU cores with the AVX application would outperform that GPU.


. . Or use Petri's approach and set the CPU crunching to the physical CPU cores and keep the HT'd cores for supporting the GPU with optimised settings. I think that would be very productive for him and leave the machine very usable for operator activity.

Stephen

:)
ID: 1859682 · Report as offensive     Reply Quote
Kets_One

Send message
Joined: 5 Jun 03
Posts: 9
Credit: 972,549
RAC: 2
Netherlands
Message 1859689 - Posted: 5 Apr 2017, 21:05:56 UTC - in response to Message 1859682.  

That sounds like a good way forward, Stephen.
The question in my mind is now: what are tyhose optimized settings for both the GPUs in my work and home computers?
ID: 1859689 · Report as offensive     Reply Quote
Stephen "Heretic"Project Donor
Volunteer tester
Avatar

Send message
Joined: 20 Sep 12
Posts: 2638
Credit: 48,865,133
RAC: 138,286
Australia
Message 1859715 - Posted: 6 Apr 2017, 0:41:56 UTC - in response to Message 1859689.  
Last modified: 6 Apr 2017, 1:34:20 UTC

That sounds like a good way forward, Stephen.
The question in my mind is now: what are those optimized settings for both the GPUs in my work and home computers?


. . Hi Kets,

. . First thing I would do, for your home computer, is rerun the installer for Lunatics Beta6. The installer will offer you AVX as the default for your CPU crunching and that is probably the best choice. For your GPU crunching SoG r3557 is pretty good and very much at the current level, but you might consider CUDA50 for the GPU in that machine as it might be a bit quicker. If you are using the default Boinc Manager then go to Options/Computing-Preferences/Computing tab and set Use at most - 50% of CPUs, and Use at most 100% of CPU time.

. . You will also need to go to your Seti account page and set the preferences there to send both CPU and GPU work.

. . Do you have a system monitor? I use Afterburner as my monitor and I like it very much. It allows me to have a monitor screen in the background that I can consult at any time. That way I can see at a glance how well the GPU is performing with the settings I use. Running one task at a time on your GPU I would see how well it utilises the GPU resources. If around the 90% mark or better I would run with that, if less than 85% I would think about changing to two tasks at a time. If you have choose CUDA50 it will only use part of one CPU HT thread whether you are running singles or doubles. If you stay with SoG it will by default use one CPU thread for each task running. With two HT threads available that is not a problem when running singles but I would stay away from running doubles unless you want to turn sleep on. If you stay with SoG there are a whole lot more settings to tweak to get the most out of it, and with the right settings it should do quite well running singles on your GPU. I do think though that it will be most productive running CUDA50 as doubles and taking the hit that comes comes when running GBT WUs. This leads us to a whole other technique that can overcome this problem to some degree. Take a peek at MrKevvy's message forum about moving CPU tasks to the GPU and vice versa.

. . The tricky part, and you may need to ask Petri for specifics, is how to lock the CPU crunching to the non HT threads. He left a message giving some details but I cannot remember which forum it is in.

. . For tweaking SoG there should be a .txt file in your Seti project folder called mb_cmdline_win_x86_SSE3_OpenCL_NV_SoG.txt in which you can add command line parameters to optimise the way SoG runs. If you want to run doubles on your system the first thing I would recommend is -use_sleep
. . I hope that helps. With the work computer that is another thing. With all the talk about the problems using work equipment I will stay away from that one.

Stephen

:)
ID: 1859715 · Report as offensive     Reply Quote
Stephen "Heretic"Project Donor
Volunteer tester
Avatar

Send message
Joined: 20 Sep 12
Posts: 2638
Credit: 48,865,133
RAC: 138,286
Australia
Message 1859725 - Posted: 6 Apr 2017, 1:50:53 UTC
Last modified: 6 Apr 2017, 2:22:42 UTC

. . @ Kets

. . I was editing and ran out of time so here is the edit.

. . For tweaking SoG there should be an empty .txt file in your Seti project folder called mb_cmdline_win_x86_SSE3_OpenCL_NV_SoG.txt in which you can add command line parameters to optimise the way SoG runs. If it is not there create it. If you want to run doubles on your system the first thing I would recommend is -use_sleep. If there are lag problems running with the singles config increase the iterations by about 50% and see if that helps.

Singles :-
-high_prec_timer -sbs 1024 -period_iterations_num 20 -tt 200

Doubles :- {not recommended}
-high_prec_timer -use_sleep -sbs 512 -period_iterations_num 30 -tt 120

. . I hope that helps. With the work computer that is another thing. With all the talk about the problems using work equipment I will stay away from that one.
ID: 1859725 · Report as offensive     Reply Quote
Stephen "Heretic"Project Donor
Volunteer tester
Avatar

Send message
Joined: 20 Sep 12
Posts: 2638
Credit: 48,865,133
RAC: 138,286
Australia
Message 1859727 - Posted: 6 Apr 2017, 1:57:39 UTC - in response to Message 1859724.  

..


. . Now that is enigmatic! :)

Stephen

:)
ID: 1859727 · Report as offensive     Reply Quote
Profile Keith Myers
Volunteer tester
Avatar

Send message
Joined: 29 Apr 01
Posts: 2452
Credit: 185,903,012
RAC: 367,674
United States
Message 1859733 - Posted: 6 Apr 2017, 2:48:25 UTC - in response to Message 1859715.  
Last modified: 6 Apr 2017, 3:01:37 UTC


. . The tricky part, and you may need to ask Petri for specifics, is how to lock the CPU crunching to the non HT threads. He left a message giving some details but I cannot remember which forum it is in.

You can use Task Manager > Processes to alter the task affinity for an application. But that is just for that one instance the application is running. Probably the easiest is to run ProcessLasso (Free) and set the affinity to Real cores for the application you want to run on physical cores and not HT or virtual cores. That applies to all instances the app is run.

I have the MB8_win_x64_AVX_VS2010_r3330.exe locked to physical cores in ProcessLasso.


Seti@Home classic workunits:20,676 CPU time:74,226 hours
ID: 1859733 · Report as offensive     Reply Quote
EdwardPF
Volunteer tester

Send message
Joined: 26 Jul 99
Posts: 341
Credit: 129,503,769
RAC: 84,621
United States
Message 1859793 - Posted: 6 Apr 2017, 12:29:36 UTC - in response to Message 1859733.  
Last modified: 6 Apr 2017, 12:33:24 UTC

You can use Task Manager > Processes to alter the task affinity for
... BOINC.EXE. All of its child processes will retain the affinity of that BOINC.EXE process ... I.E. any MB8_win_x86... processes.

Ed F

(you need to do this once per boot ... or at S@H startup time with a "/affinity 55" for cpu's 0:2:4:6 ... or whatever).
ID: 1859793 · Report as offensive     Reply Quote
Kets_One

Send message
Joined: 5 Jun 03
Posts: 9
Credit: 972,549
RAC: 2
Netherlands
Message 1859798 - Posted: 6 Apr 2017, 12:48:33 UTC

Guys,

Many thanks for all your responses.
Im currently trying out the 'singles' setting.

Kets
ID: 1859798 · Report as offensive     Reply Quote
FloydHung

Send message
Joined: 22 Jul 99
Posts: 7
Credit: 1,060,648
RAC: 1,292
Taiwan
Message 1860703 - Posted: 10 Apr 2017, 7:44:29 UTC - in response to Message 1859578.  

Thanks for your help. It works after adding it whitelists.
ID: 1860703 · Report as offensive     Reply Quote
Kets_One

Send message
Joined: 5 Jun 03
Posts: 9
Credit: 972,549
RAC: 2
Netherlands
Message 1861476 - Posted: 14 Apr 2017, 18:09:14 UTC

Hey Guys,

It has been a few days since my last message and i would like to share my experiences.
I decided to ditch the work computer and not to use it for SETI work. :)

For my home computer i decided to stick with SoG for now and see how these optimizations work out.
I installed the Lunatics beta 0.45 beta 6 and the Nvidia r_3584, also i decided to crunch with CPU only as these appear to have the most opportunities for optimizations.
CPU utilization looks around 99% (it runs HOT, also when using "-use_sleep").

I am using the following optimizations: -high_prec_timer -sbs 1024 -period_iterations_num 20 -tt 200

After a few days running i appear to get differently sized WUs now, they seem a lot smaller with sizes varying between ~9,000 GFLOPs and ~90,000 GFLOPs, where previous i was getting bigger ones regularly (up to ~200,000). Is this normal? Also, im getting only Seti@home version 8.20 WUs, no 8.22 i was getting earlier before any optimizations.

Now im getting the following results:
WU 5641503303 - runtime 15 mins (wallclock)- 45.29 points

This still seems a bit low to me. Am i wrong?
ID: 1861476 · Report as offensive     Reply Quote
Profile Keith Myers
Volunteer tester
Avatar

Send message
Joined: 29 Apr 01
Posts: 2452
Credit: 185,903,012
RAC: 367,674
United States
Message 1861481 - Posted: 14 Apr 2017, 19:01:48 UTC - in response to Message 1861476.  

The difference in version numbers 8.22 vice 8.20 is just because the Lunatics installer used the 8.20 version number in its generated app_info. The 8.22 version number came from the stock app. The Lunatics Beta-06 installer is one app version behind the stock app. No real difference in performance between the apps. The tasks are all the same. No difference. And regarding the low credit awarded for your referenced task.... well welcome to CreditScrew. Nothing you can do about it and is totally normal. The credit awarded has no real correlation to the flops required to process the task now. Just accept it.
Seti@Home classic workunits:20,676 CPU time:74,226 hours
ID: 1861481 · Report as offensive     Reply Quote
Kets_One

Send message
Joined: 5 Jun 03
Posts: 9
Credit: 972,549
RAC: 2
Netherlands
Message 1861486 - Posted: 14 Apr 2017, 20:18:22 UTC - in response to Message 1861476.  
Last modified: 14 Apr 2017, 20:18:47 UTC

Correction, i meant to say i am using GPU-only:


I installed the Lunatics beta 0.45 beta 6 and the Nvidia r_3584, also i decided to crunch with GPU only as these appear to have the most opportunities for optimizations.
GPU utilization looks around 99% (it runs HOT, also when using "-use_sleep").

ID: 1861486 · Report as offensive     Reply Quote
Stephen "Heretic"Project Donor
Volunteer tester
Avatar

Send message
Joined: 20 Sep 12
Posts: 2638
Credit: 48,865,133
RAC: 138,286
Australia
Message 1861495 - Posted: 14 Apr 2017, 21:50:01 UTC - in response to Message 1861476.  
Last modified: 14 Apr 2017, 22:10:57 UTC

Hey Guys,

It has been a few days since my last message and i would like to share my experiences.
I decided to ditch the work computer and not to use it for SETI work. :)

Probably a wise and cautious decision :)


For my home computer i decided to stick with SoG for now and see how these optimizations work out.
I installed the Lunatics beta 0.45 beta 6 and the Nvidia r_3584, also i decided to crunch with CPU only as these appear to have the most opportunities for optimizations.
CPU utilization looks around 99% (it runs HOT, also when using "-use_sleep").

. . Lunatics Beta6 only incorporates r3557 which is V8.20 not v8.22, but this is not a serious issue. The difference is about inconclusive results and is only an issue if you were/are experiencing a high rate of inconclusives with r3557, which I am not. Hopefully you will not either. Did you download Afterburner? It can not only show you exactly what your system utilisation is but allow you to create a custom fan profile to keep the temps down.


I am using the following optimizations: -high_prec_timer -sbs 1024 -period_iterations_num 20 -tt 200

After a few days running i appear to get differently sized WUs now, they seem a lot smaller with sizes varying between ~9,000 GFLOPs and ~90,000 GFLOPs, where previous i was getting bigger ones regularly (up to ~200,000). Is this normal? Also, im getting only Seti@home version 8.20 WUs, no 8.22 i was getting earlier before any optimizations.

Now im getting the following results:
WU 5641503303 - runtime 15 mins (wallclock)- 45.29 points

This still seems a bit low to me. Am i wrong?


. . They are what I call halflings and most people call shorties. They are Arecibo WUs from data that has a high AR or angular ratio. This means they are scanning more quickly and covering a much bigger part of the sky. This causes them to register slightly less power for individual signal sources and so only have the "brighter" objects to deal with. They crunch much more quickly than normal AR tasks. They take about half the time to process (which is why I call them halflings). Because of the lower load and shorter processing times CreditScrew only assigns them about 40 points.

. . Looking at your results it seems you are using -sbs 2048, it never can work at that as you only have 2048K of video ram. Never set it higher than about half of your video ram or less. It is defaulting to 512 so probably that is what you should set it to.

. . Otherwise it sounds to me like you are on the right track.

Stephen

:)
ID: 1861495 · Report as offensive     Reply Quote
Profile Wiggo "Socialist"
Avatar

Send message
Joined: 24 Jan 00
Posts: 12615
Credit: 169,679,224
RAC: 87,266
Australia
Message 1861502 - Posted: 14 Apr 2017, 22:31:04 UTC

I installed the Lunatics beta 0.45 beta 6 and the Nvidia r_3584, also i decided to crunch with GPU only as these appear to have the most opportunities for optimizations.
GPU utilization looks around 99% (it runs HOT, also when using "-use_sleep").

I don't know why people have missed reading this bit or looking at the completed results stating that r3584 is being used. :-(

Try lifting your laptop by using something (sticks, lego blocks, cooling pad, etc,..) to let more cool air to circulate under it as this usually helps temps. ;-)

Cheers.
ID: 1861502 · Report as offensive     Reply Quote
Profile Keith Myers
Volunteer tester
Avatar

Send message
Joined: 29 Apr 01
Posts: 2452
Credit: 185,903,012
RAC: 367,674
United States
Message 1861511 - Posted: 15 Apr 2017, 0:36:53 UTC - in response to Message 1861502.  

I installed the Lunatics beta 0.45 beta 6 and the Nvidia r_3584, also i decided to crunch with GPU only as these appear to have the most opportunities for optimizations.
GPU utilization looks around 99% (it runs HOT, also when using "-use_sleep").

I don't know why people have missed reading this bit or looking at the completed results stating that r3584 is being used. :-(

Try lifting your laptop by using something (sticks, lego blocks, cooling pad, etc,..) to let more cool air to circulate under it as this usually helps temps. ;-)

Cheers.


Well unless Richard has slipstreamed r3584 into the Lunatics 0.45 Beta-06 without telling anyone, the installer will install the r3557 SoG app. And since I believe we all thought the OP was a noobie, that it was beyond considering that he had hand edited the r3584 SoG app into his app_info. Obviously some assumptions were made. I see the r3557 app being used on one computer, I believe the laptop, and the other computer IS using the r3584 SoG app.

So has anyone run the Lunatics 0.45 Beta-06 installer lately and discovered that in fact it installs the r3584 SoG app?
Seti@Home classic workunits:20,676 CPU time:74,226 hours
ID: 1861511 · Report as offensive     Reply Quote
Stephen "Heretic"Project Donor
Volunteer tester
Avatar

Send message
Joined: 20 Sep 12
Posts: 2638
Credit: 48,865,133
RAC: 138,286
Australia
Message 1861512 - Posted: 15 Apr 2017, 0:44:44 UTC - in response to Message 1861502.  
Last modified: 15 Apr 2017, 1:23:24 UTC

I installed the Lunatics beta 0.45 beta 6 and the Nvidia r_3584, also i decided to crunch with GPU only as these appear to have the most opportunities for optimizations.
GPU utilization looks around 99% (it runs HOT, also when using "-use_sleep").

I don't know why people have missed reading this bit or looking at the completed results stating that r3584 is being used. :-(

Try lifting your laptop by using something (sticks, lego blocks, cooling pad, etc,..) to let more cool air to circulate under it as this usually helps temps. ;-)

Cheers.


. . Hi Wiggo, perhaps you didn't look at enough of his results. r3557 is being used. But while there are some results showing r3584 as well, I considered that they would have been tasks left over from before Lunatics and were still labelled as V8.22 thus triggering the already installed r3584 app for those tasks. That and the fact he questions why he is now getting V8.20 tasks and NOT V8.22 tasks as he had been previously makes my conclusion pretty much inevitable.

. . But maybe he did jump right on in and edit his app_info.xml as well .... do you want to ask him?

[edit] After writing the above I had another look at his results, and tasks downloaded in the last 24 hours, after lunatics, are showing r3584. So he must have edited app_info.xml after all, and his question about v8.20 tasks may have been premature, or maybe they are simply still being labelled as V8.20??. Maybe I should ask him for some tips? :)

Stephen

:)
ID: 1861512 · Report as offensive     Reply Quote
Profile Keith Myers
Volunteer tester
Avatar

Send message
Joined: 29 Apr 01
Posts: 2452
Credit: 185,903,012
RAC: 367,674
United States
Message 1861516 - Posted: 15 Apr 2017, 0:55:04 UTC - in response to Message 1861512.  

But the latest GPU tasks were run under the Anonymous platform. Having the r3584 app in the stderr.txt file only would be possible if he started the task as stock and then midway through transitioned to the Anonymous platform to finish it.
Seti@Home classic workunits:20,676 CPU time:74,226 hours
ID: 1861516 · Report as offensive     Reply Quote
Stephen "Heretic"Project Donor
Volunteer tester
Avatar

Send message
Joined: 20 Sep 12
Posts: 2638
Credit: 48,865,133
RAC: 138,286
Australia
Message 1861517 - Posted: 15 Apr 2017, 1:09:56 UTC - in response to Message 1861516.  

But the latest GPU tasks were run under the Anonymous platform. Having the r3584 app in the stderr.txt file only would be possible if he started the task as stock and then midway through transitioned to the Anonymous platform to finish it.


. . When running the Lunatics installer any WUs already in your cache retain their name/description in the client_state.xml. So there would still have been V8.22 tasks in his queue after lunatics was running. I may be wrong but I am of the opinion that because r3584 is already in his app_info.xml when those tasks started they triggered that app not r3557 while otherwise showing the lunatics environment. But there are many other results showing lunatics and r3557. I also was surprised by the r3584 results so I looked at several others before reaching a conclusion. One thing that struck me but may just be random chance is that the results showing the r3584 app all have shorter run times than those showing r3557. I am now sorely tempted to risk catastrophe and edit app_info.xml to manually install r3584.

Stephen

:)
ID: 1861517 · Report as offensive     Reply Quote
Previous · 1 · 2 · 3 · 4 · 5 · 6 . . . 7 · Next

Message boards : Number crunching : SETI@home v8.22 Windows GPU applications support thread


 
©2017 University of California
 
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.