Multiple S@H threads per GPU

Questions and Answers : Windows : Multiple S@H threads per GPU
Message board moderation

To post messages, you must log in.

AuthorMessage
Mark Seeger

Send message
Joined: 16 May 99
Posts: 47
Credit: 16,558,494
RAC: 116
United States
Message 1947990 - Posted: 5 Aug 2018, 6:15:07 UTC

Currently I am only able to run one S@H work unit per GPU. Does anyone know how to run multiple threads per GPU?

For instance, my nvidia Titan Xp are running at 1/4 power draw while crunching only one work unit at a time (albeit very fast: 3-4 minutes per work unit).

Thank you,

Mark
ID: 1947990 · Report as offensive
Profile Tom M
Volunteer tester

Send message
Joined: 28 Nov 02
Posts: 5124
Credit: 276,046,078
RAC: 462
Message 1948688 - Posted: 9 Aug 2018, 22:33:09 UTC - in response to Message 1947990.  
Last modified: 9 Aug 2018, 22:37:35 UTC

Create an "app_config.xml" file. Place in in the seti directory under the hidden progamdata .
Add this to file. This will assign 3 tasks to the gpu. It will also use 1 cpu core per task. This "should" maximize the speed of processing.

Contents of app_config.xml file.
<app_config>
<app>
<name>setiathome_v8</name>
<gpu_versions>
<gpu_usage>0.33</gpu_usage>
<cpu_usage>1.0</cpu_usage>
</gpu_versions>
</app>
</app_config>

I am assuming you already have a MB*SOG.txt command line file in place. It probably looks something like this: -sbs 256 -spike_fft_thresh 4096 -tune 1 64 1 4 2 64 1 4 -oclfft_tune_gr 256 -oclfft_tune_lr 16 -oclfft_tune_wg 256 -oclfft_tune_ls 512 -oclfft_tune_bn 64 -oclfft_tune_cw 64

HTH,
Tom
A proud member of the OFA (Old Farts Association).
ID: 1948688 · Report as offensive
Profile eng4hire

Send message
Joined: 25 Dec 10
Posts: 490
Credit: 34,584,466
RAC: 0
United States
Message 1953304 - Posted: 2 Sep 2018, 0:25:19 UTC - in response to Message 1947990.  

TOM,

Q1 - What is the relationship of CPU cores to GPU tasks or threads?
Q2 - Why is it that my current system uses a fractional CPU core for one GPU card?

My system has multiple NVIDIA Quadro 4000 cards but as Mark points out, each GPU of mine is only crunching one task at a time.
Unlike Marks environment and not familiar with his video card, mine all seem to be running at MAX PWR and TEMP I might add.

Q3 - If I follow your instructions to Mark and do the same thing, will I be "overloading" my GPU's?
Q4 - You mention that Mark also needs this other file. I assume both are required to make this happen?
Q5 - How many tasks are feasible to run on a GPU?

Thanks,

Dave
ID: 1953304 · Report as offensive
Profile Tom M
Volunteer tester

Send message
Joined: 28 Nov 02
Posts: 5124
Credit: 276,046,078
RAC: 462
Message 1953336 - Posted: 2 Sep 2018, 3:03:05 UTC - in response to Message 1953304.  

Dave,
Until you reply in PM I won't be able to "see" the reply I sent to your PM of the above public message. So would you please copy my reply here?

Thank you,
Tom
A proud member of the OFA (Old Farts Association).
ID: 1953336 · Report as offensive
rob smith Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer moderator
Volunteer tester

Send message
Joined: 7 Mar 03
Posts: 21309
Credit: 416,307,556
RAC: 380
United Kingdom
Message 1953357 - Posted: 2 Sep 2018, 8:24:54 UTC - in response to Message 1953304.  

TOM,

Q1 - What is the relationship of CPU cores to GPU tasks or threads?
Q2 - Why is it that my current system uses a fractional CPU core for one GPU card?

My system has multiple NVIDIA Quadro 4000 cards but as Mark points out, each GPU of mine is only crunching one task at a time.
Unlike Marks environment and not familiar with his video card, mine all seem to be running at MAX PWR and TEMP I might add.

Q3 - If I follow your instructions to Mark and do the same thing, will I be "overloading" my GPU's?
Q4 - You mention that Mark also needs this other file. I assume both are required to make this happen?
Q5 - How many tasks are feasible to run on a GPU?

Thanks,

Dave

Some good questions, and hopefully some constructive answers:
Q1 - Any GPU task needs a bit of CPU support, for the "stock" (non-SoG) application this is quite small, less than 10%
Q2 - When a task is deployed by the server a default amount of CPU and GPU effort is assigned. It is possible to over-ride these using configuration files, however care must be exercised - do not exceed 1 CPU per task as the current crop of applications can only use one CPU core/thread.
Q3 - Not really, as the GPU will simply start to swap between tasks and run much slower if you exceed its potential.
Q5 - That entirely depends on what Mark is trying to do.
Q6 - That depends on what application you are running, what operating system, what the CPU is, how much CPU memory you have, and no doubt a few other things. For example, if you are running Linux and the "special" application you can only run one task per GPU, but if you are running the "SoG" application under Windows then you can run more than one. You have to try and see what works best in your situation.
Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?
ID: 1953357 · Report as offensive
Profile Tom M
Volunteer tester

Send message
Joined: 28 Nov 02
Posts: 5124
Credit: 276,046,078
RAC: 462
Message 1953384 - Posted: 2 Sep 2018, 13:22:08 UTC - in response to Message 1953357.  

Thank you Bob.

Dave,
You probably would have to use the command for "use all gpu's" unless the gpu's on the Mining Rig were all the same make/model.

Tom
A proud member of the OFA (Old Farts Association).
ID: 1953384 · Report as offensive
rob smith Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer moderator
Volunteer tester

Send message
Joined: 7 Mar 03
Posts: 21309
Credit: 416,307,556
RAC: 380
United Kingdom
Message 1953409 - Posted: 2 Sep 2018, 15:13:30 UTC - in response to Message 1947990.  

Dave,
You cannot take the fraction of power drawn as a metric for the fraction of a GPU in use, as GPUs have whole areas of their silicon that are not heavily used by the CUDA or SoG applications. These unused parts are to do with graphic processing and driving.
Depending on the GPU it is possible to get a metric for the fraction of the computational components and RAM that are in use, which are a better indication of GPU loading under computational loads. For many GPUs tools like "GPU-Z" will return these figures as an average over a fairly short time frame.
Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?
ID: 1953409 · Report as offensive
Mark Seeger

Send message
Joined: 16 May 99
Posts: 47
Credit: 16,558,494
RAC: 116
United States
Message 1955311 - Posted: 14 Sep 2018, 6:36:53 UTC - in response to Message 1948688.  

Hi Tom,

I'm catching up on this thread (forgive my delay).

Question: where are these files located (or, to be located) on Windoze 10? My PC with the dual Titan XPs is running Windows and Im not familiar with the OS file locations (I'm a Mac user :). Any pointers will be greatly appreciated. Thank you,

Mark
ID: 1955311 · Report as offensive
rob smith Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer moderator
Volunteer tester

Send message
Joined: 7 Mar 03
Posts: 21309
Credit: 416,307,556
RAC: 380
United Kingdom
Message 1955326 - Posted: 14 Sep 2018, 12:57:11 UTC

By default the files are held in a hidden folder c:\ProgramData\BOINC\projects\project_name
Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?
ID: 1955326 · Report as offensive
Mark Seeger

Send message
Joined: 16 May 99
Posts: 47
Credit: 16,558,494
RAC: 116
United States
Message 1955350 - Posted: 14 Sep 2018, 16:46:15 UTC - in response to Message 1955326.  

Thank you Bob.

I created the xml file Tom referenced, and I copied/pasted his lines exactly as shown, rebooted the computer, however SETI is still running as before (one task per GPU).

Should I edit the XML text? I copied it exactly as Tom suggested above, and placed it in the project data folder.

Thank you,

Mark
ID: 1955350 · Report as offensive
Mark Seeger

Send message
Joined: 16 May 99
Posts: 47
Credit: 16,558,494
RAC: 116
United States
Message 1955351 - Posted: 14 Sep 2018, 16:55:09 UTC - in response to Message 1955350.  

Got it working. Sorry, I didnt notice that Windoze 10 kept a file extension. All set :)
ID: 1955351 · Report as offensive
Profile Tom M
Volunteer tester

Send message
Joined: 28 Nov 02
Posts: 5124
Credit: 276,046,078
RAC: 462
Message 1955425 - Posted: 14 Sep 2018, 23:11:20 UTC - in response to Message 1955351.  

Hurray!!!
A proud member of the OFA (Old Farts Association).
ID: 1955425 · Report as offensive
Profile Tom M
Volunteer tester

Send message
Joined: 28 Nov 02
Posts: 5124
Credit: 276,046,078
RAC: 462
Message 1955507 - Posted: 15 Sep 2018, 13:30:55 UTC - in response to Message 1955351.  

Got it working. Sorry, I didnt notice that Windoze 10 kept a file extension. All set :)


I have been caught by that "several" times :)

What kind of time are they taking now? Previously you should have had pretty low times (well below 7 minutes / task). Based on some hearsay, that time might increase upwards to 30% higher but you get three times the production.

Tom
A proud member of the OFA (Old Farts Association).
ID: 1955507 · Report as offensive

Questions and Answers : Windows : Multiple S@H threads per GPU


 
©2023 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.