CPU utalization

Message boards : Number crunching : CPU utalization
Message board moderation

To post messages, you must log in.

AuthorMessage
Ryan Munro

Send message
Joined: 5 Feb 06
Posts: 63
Credit: 17,099,460
RAC: 10,172
United Kingdom
Message 1981290 - Posted: 20 Feb 2019, 14:45:17 UTC

I have a 7820x and a Titan Xp running in my rig, always been taught to leave a CPU free for that GPU.
So if I set it to only run 14 CPU tasks the CPU is maxed out completely in task manager, I have to drop it back to 11 CPU task to see the CPU utilization drop just under 100%.
Is that right?, If anything above 11 CPU tasks maxes out the CPU I assume that 14 would be choking the GPU?

In my head, I would assume 14 tasks and 1 GPU should have the CPU running about 95%?

Any technical reason why running less than 16 tasks on a chip that can handle 16 threads would still max out the CPU?
ID: 1981290 · Report as offensive
TBar
Volunteer tester

Send message
Joined: 22 May 99
Posts: 4864
Credit: 590,595,100
RAC: 1,387,976
United States
Message 1981299 - Posted: 20 Feb 2019, 15:46:54 UTC - in response to Message 1981290.  

It depends on what else is running on the system. On my 12 GPU system I've noted just running nvidia-settings is using a full CPU core.
To see what your GPU & System needs, in the BOINC Manager Advanced View, Suspend all the CPU tasks and then note the CPU usage with just the GPU and System running.
ID: 1981299 · Report as offensive
mmonnin
Volunteer tester

Send message
Joined: 8 Jun 17
Posts: 58
Credit: 10,176,409
RAC: 0
United States
Message 1981302 - Posted: 20 Feb 2019, 16:01:07 UTC

I've found it also depends on the OS. Linux seems to be better at giving the GPU enough CPU power to keep the GPU fully utilized. Windows (at least Win7) will end up choking the GPU for CPU power and GPU utilization will drop or be very up and down. I run my Linux (debian) systems at max CPU tasks along with multiple GPU tasks. In Win7 I leave a thread open for the GPU, even if the GPU uses very little CPU like OpenCL ATI tasks.
ID: 1981302 · Report as offensive
Profile Siran d'Vel'nahr
Volunteer tester
Avatar

Send message
Joined: 23 May 99
Posts: 7052
Credit: 25,116,001
RAC: 106,478
United States
Message 1981303 - Posted: 20 Feb 2019, 16:06:35 UTC - in response to Message 1981299.  

It depends on what else is running on the system. On my 12 GPU system I've noted just running nvidia-settings is using a full CPU core.
To see what your GPU & System needs, in the BOINC Manager Advanced View, Suspend all the CPU tasks and then note the CPU usage with just the GPU and System running.

Hi TBar,

I have yet to find a way to just suspend the CPU and not the GPU. I can suspend the GPU and not the CPU, but not the other way around. Is there a work around to do that?

Have a great day! :)

Siran
CAPT Siran d'Vel'nahr XO - L L & P _\\//
USS Vre'kasht NCC-33187

Windows 10 OS? "What a piece of junk!" - L. Skywalker

ID: 1981303 · Report as offensive
TBar
Volunteer tester

Send message
Joined: 22 May 99
Posts: 4864
Credit: 590,595,100
RAC: 1,387,976
United States
Message 1981306 - Posted: 20 Feb 2019, 16:19:33 UTC - in response to Message 1981303.  
Last modified: 20 Feb 2019, 16:30:50 UTC

Yes, in the BOINC Manager Advanced view Tasks Tab, Select All the CPU tasks and hit the Suspend button. It's easier if you sort by Application and use the Shift key.

mmonnin is using old version of the App that has the -pfl 64 setting. That version was found to have problems with that setting.
Anyone running that version needs tp Update to the Current version at C.A, http://www.arkayn.us/lunatics/BOINC.7z
If you see this in the stderr_txt, you should Update;
Using default pulse Fft limit (-pfl 64)
setiathome v8 enhanced x41p_V0.97b1, Cuda 9.10 special
Many people have found with the Lower End GPUs in Linux you still see a Large GPU Slow Down if you don't Reserve a CPU core for the GPU. I certainly do. Using -nobs helps the GPU, but then the CPU tasks Slow Down. It's best to reserve at least 1 CPU core even in Linux.
ID: 1981306 · Report as offensive
Profile Siran d'Vel'nahr
Volunteer tester
Avatar

Send message
Joined: 23 May 99
Posts: 7052
Credit: 25,116,001
RAC: 106,478
United States
Message 1981308 - Posted: 20 Feb 2019, 16:27:48 UTC - in response to Message 1981306.  

Yes, in the BOINC Manager Advanced view Tasks Tab, Select All the CPU tasks and hit the Suspend button. It's easier if you sort by Application and use the Shift key.

mmonnin is using old version of the App that has the -pfl 64 setting. That version was found to have problems with that setting.
Anyone running that version needs tp Update to the Current version at C.A, http://www.arkayn.us/lunatics/BOINC.7z
If you see this in the stderr_txt, you should Update;
Using default pulse Fft limit (-pfl 64)
setiathome v8 enhanced x41p_V0.97b1, Cuda 9.10 special

Hi TBar,

Doh!!! Oh DUH!!! ;) I was trying to use the Options window. Thanks! :)

Have a great day! :)

Siran
CAPT Siran d'Vel'nahr XO - L L & P _\\//
USS Vre'kasht NCC-33187

Windows 10 OS? "What a piece of junk!" - L. Skywalker

ID: 1981308 · Report as offensive
juan BFP Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 16 Mar 07
Posts: 8130
Credit: 497,213,650
RAC: 390,106
Panama
Message 1981317 - Posted: 20 Feb 2019, 16:58:54 UTC - in response to Message 1981290.  

Any technical reason why running less than 16 tasks on a chip that can handle 16 threads would still max out the CPU?

Technical reason no, practically yes.

Each CPU has a sweet spot in the number of thread it could run to produce more daily output. Something like less is best. Weird no?

In my 6850 (12 threads) running 4 GPU WU, the sweet spot is 4 CPU WU, that leaves up to 4 Thread free i know but if you sum the daily CPU production with 4 WU and 6 or even 8 WU you will notice is about the same. And of course running just 4 the CPU runs a lot cooler than 6 or 8 and uses less electric power.

Keith do the same tests on his hosts and find something very similar, run more WU not always makes the host produce more, so you need to test in your particular system. As always YMMV.

Explain why that happening is above my paid grade. What i know it's something directly related on what is running and how you use the host.
ID: 1981317 · Report as offensive
mmonnin
Volunteer tester

Send message
Joined: 8 Jun 17
Posts: 58
Credit: 10,176,409
RAC: 0
United States
Message 1981329 - Posted: 20 Feb 2019, 17:43:45 UTC - in response to Message 1981306.  

Yes, in the BOINC Manager Advanced view Tasks Tab, Select All the CPU tasks and hit the Suspend button. It's easier if you sort by Application and use the Shift key.

mmonnin is using old version of the App that has the -pfl 64 setting. That version was found to have problems with that setting.
Anyone running that version needs tp Update to the Current version at C.A, http://www.arkayn.us/lunatics/BOINC.7z
If you see this in the stderr_txt, you should Update;
Using default pulse Fft limit (-pfl 64)
setiathome v8 enhanced x41p_V0.97b1, Cuda 9.10 special
Many people have found with the Lower End GPUs in Linux you still see a Large GPU Slow Down if you don't Reserve a CPU core for the GPU. I certainly do. Using -nobs helps the GPU, but then the CPU tasks Slow Down. It's best to reserve at least 1 CPU core even in Linux.


No, I'm not. I see this in my output. I have changed no settings/command lines.
setiathome v8 enhanced x41p_V0.97b1, Cuda 9.10 special

If there's any tasks you'll find its the 9.2 version which have 100% CPU utilization so I switched back to 9.1

Any my post was BOINC project agnostic.
ID: 1981329 · Report as offensive
TBar
Volunteer tester

Send message
Joined: 22 May 99
Posts: 4864
Credit: 590,595,100
RAC: 1,387,976
United States
Message 1981332 - Posted: 20 Feb 2019, 17:54:31 UTC - in response to Message 1981329.  

I see it in all the current tasks, Received: 20 Feb 2019, 17:44:03 UTC
Using default pulse Fft limit (-pfl 64)

If you don't want to change Apps, then at least add the cmdline -pfl 512 to your current App.
I think you will find the CUDA 9.0 App in the link is faster than the 9.1 App you are using, and it uses pfl 512.
ID: 1981332 · Report as offensive
mmonnin
Volunteer tester

Send message
Joined: 8 Jun 17
Posts: 58
Credit: 10,176,409
RAC: 0
United States
Message 1981370 - Posted: 20 Feb 2019, 21:34:29 UTC - in response to Message 1981332.  

I see it in all the current tasks, Received: 20 Feb 2019, 17:44:03 UTC
Using default pulse Fft limit (-pfl 64)

If you don't want to change Apps, then at least add the cmdline -pfl 512 to your current App.
I think you will find the CUDA 9.0 App in the link is faster than the 9.1 App you are using, and it uses pfl 512.


The maxwell app is faster than the pascal app for a 1080Ti? I found 9.2 to be virtually the same as 9.1 without modifications to either. CPU usage was much worse so I went back to 9.1.
ID: 1981370 · Report as offensive
TBar
Volunteer tester

Send message
Joined: 22 May 99
Posts: 4864
Credit: 590,595,100
RAC: 1,387,976
United States
Message 1981396 - Posted: 20 Feb 2019, 23:57:32 UTC - in response to Message 1981370.  
Last modified: 21 Feb 2019, 0:03:58 UTC

Any CUDA version 9.0 and above supports Pascal, as long as the App was compiled with the CC=6.1 option. The only thing that makes it a 'Maxwell' App was if it was also compiled with CC=5.0 & 5.2. The 9.2 App wasn't compiled with 5&5.2, the other Apps were. In my tests the CUDA 9.0 App is just as fast, or faster, as any other 9.x App with Pascal. 9.1 & 9.2 have Native support for Volta, CUDA 10.0 has Native support for Turing. If all you have are Pascal GPUs, CUDA 9.0 will work just as well as any other 9.x App. My tests also show the CUDA 10.0 App to be a couple of seconds faster than the CUDA 9 App, However, you will need a CUDA 10 driver to use the CUDA 10 App. All of the Newer Apps use pfl 512, which by default makes them better than the CUDA 9.1 App you are using.

BTW, the setting -nobs controls how much CPU is used by the GPU. Add/Remove the -nobs setting from the app_info to use more or less CPU. It's in the ReadMe.
ID: 1981396 · Report as offensive

Message boards : Number crunching : CPU utalization


 
©2019 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.