(x.xx CPUs, 1 CUDA)


log in

Advanced search

Message boards : Number crunching : (x.xx CPUs, 1 CUDA)

1 · 2 · Next
Author Message
Profile Chirag Patel
Volunteer tester
Avatar
Send message
Joined: 13 Sep 05
Posts: 48
Credit: 7,894,510
RAC: 4,297
India
Message 941296 - Posted: 19 Oct 2009, 8:45:56 UTC

So, I've noticed this line in my CUDA tasks, that says "Running (0.04 CPUs, 1 CUDA)". I assume this means that 4% of one CPU is being used to supply work to the GPU. Is that enough?

I have a feeling that the GPU isn't doing as much work as it could. For example, when I watch a movie and I forget to switch off BOINC, it still runs for a minute or so before the occasional stuttering. Also, Windows 7's Aero feature (say minimising windows) is unaffected most of the time even when BOINC is running. Does this mean my GPU is idle all that time?
____________

Profile ignorance is no excuse
Avatar
Send message
Joined: 4 Oct 00
Posts: 9529
Credit: 44,433,274
RAC: 0
Korea, North
Message 941338 - Posted: 19 Oct 2009, 14:13:30 UTC - in response to Message 941296.

if you look on your preferences on the BOINC manager you see the "run GPU when computer is being used" if its unclicked and you are using your PC then no it won't be running.
____________
In a rich man's house there is no place to spit but his face.
Diogenes Of Sinope

End terrorism by building a school

Profile Chirag Patel
Volunteer tester
Avatar
Send message
Joined: 13 Sep 05
Posts: 48
Credit: 7,894,510
RAC: 4,297
India
Message 941351 - Posted: 19 Oct 2009, 15:12:00 UTC - in response to Message 941338.
Last modified: 19 Oct 2009, 15:12:41 UTC

if you look on your preferences on the BOINC manager you see the "run GPU when computer is being used" if its unclicked and you are using your PC then no it won't be running.

Yeah, I've clicked that option, so I know it's running even when I'm using my PC (also the CUDA task says 'Running' in the manager, not 'Waiting to run'). But it still hardly has any impact on the computer.
____________

Profile ignorance is no excuse
Avatar
Send message
Joined: 4 Oct 00
Posts: 9529
Credit: 44,433,274
RAC: 0
Korea, North
Message 941352 - Posted: 19 Oct 2009, 15:21:27 UTC - in response to Message 941351.

I'd be willing to bet the 8800 CUda card would cause more problems than the GTX260 the better the card the less you'll notice work being done. I notice when I run my ATI 4770 when gaming. not so much when its off.
____________
In a rich man's house there is no place to spit but his face.
Diogenes Of Sinope

End terrorism by building a school

Profile Misfit
Volunteer tester
Avatar
Send message
Joined: 21 Jun 01
Posts: 21790
Credit: 2,510,901
RAC: 0
United States
Message 941375 - Posted: 19 Oct 2009, 16:34:55 UTC - in response to Message 941296.

So, I've noticed this line in my CUDA tasks, that says "Running (0.04 CPUs, 1 CUDA)". I assume this means that 4% of one CPU is being used to supply work to the GPU. Is that enough?

If I remember correctly that has to do more with the work fetch than the CPU resources. Look at your task manager and you'll see the CPU percentage jumping around in the GPU process.
____________

Josef W. Segur
Volunteer developer
Volunteer tester
Send message
Joined: 30 Oct 99
Posts: 4202
Credit: 1,030,017
RAC: 264
United States
Message 941424 - Posted: 19 Oct 2009, 20:27:26 UTC - in response to Message 941296.

So, I've noticed this line in my CUDA tasks, that says "Running (0.04 CPUs, 1 CUDA)". I assume this means that 4% of one CPU is being used to supply work to the GPU. Is that enough?

I have a feeling that the GPU isn't doing as much work as it could. For example, when I watch a movie and I forget to switch off BOINC, it still runs for a minute or so before the occasional stuttering. Also, Windows 7's Aero feature (say minimising windows) is unaffected most of the time even when BOINC is running. Does this mean my GPU is idle all that time?

It is not a measurement, it's an estimate. For those running stock, there's a server-side estimate of how much CPU will be used, based on the host's benchmarks and the estimated speed of the GPU. For those running anonymous platform, the 0.04 is what's in the app_info.xml. It makes very little difference, the BOINC core client uses it to decide whether a CPU application can be run at the same time as a GPU application. In effect the 0.04 says 96% of the CPU is available for other work.

Actual CPU usage can be checked with Task Manager, of course. It isn't steady usage, the GPU only needs to be told what to do next after it has completed an operation.

Some users have determined that they get the best productivity by leaving one CPU core effectively idle so it can feed the GPU with no delay. That mode would probably speed up the GPU work on your host, but whether it would be enough to make up for the unused CPU is uncertain.
Joe

Richard Haselgrove
Volunteer tester
Send message
Joined: 4 Jul 99
Posts: 8375
Credit: 46,670,695
RAC: 18,746
United Kingdom
Message 941437 - Posted: 19 Oct 2009, 21:29:57 UTC - in response to Message 941424.

Some users have determined that they get the best productivity by leaving one CPU core effectively idle so it can feed the GPU with no delay. That mode would probably speed up the GPU work on your host, but whether it would be enough to make up for the unused CPU is uncertain.
Joe

I can't imagine that it would be worth sacrificing a whole CPU core to feed one measly mid-range CUDA card, but I can see that it could help for top crunchers like Simon or Vyper with 6 or 8 top-of-the-range cards to feed.

Profile Chirag Patel
Volunteer tester
Avatar
Send message
Joined: 13 Sep 05
Posts: 48
Credit: 7,894,510
RAC: 4,297
India
Message 941460 - Posted: 19 Oct 2009, 23:35:55 UTC

All right, thanks for the information everyone.

I'll play around with the one-CPU setting to see if my GPU WUs take less time. I have a Core 2 Duo P7350 (OC'd to 2.5 GHz), not the best CPU, so maybe the extra GPU usage will make up for the lack of a CPU core. And yes, CPU usage for the CUDA application jumps between 0% and 10% most of the time.
____________

Profile [seti.international] Dirk Sadowski
Volunteer tester
Avatar
Send message
Joined: 6 Apr 07
Posts: 7022
Credit: 59,232,833
RAC: 20,641
Germany
Message 941470 - Posted: 20 Oct 2009, 0:13:50 UTC
Last modified: 20 Oct 2009, 0:30:10 UTC


On my GPU cruncher, only - like the name ;-) - calculation on the GPUs.

If all stock I had (0.19 CPU, 1 CUDA)
Now with opt._CUDA_app I have no specially entry in my app_info.xml, so there is (1.00 CPU, 1 CUDA).
Because of, if not BOINC would ask continuously for new work for 4 idle CPU-Cores.

If I would crunch also on the CPU, the GPU performance would go massively down.
Only on (whole) CPU, maybe ~ 4,000 RAC.
Only on GPUs, last stable RAC as 24/7 ~ 52,000. ~ 13,000/GPU.

If CPU and GPUs simultaneously, the GPU calculation time increased ~ x3.
Because of boinc.exe (BOINC client) and System activity/peaks [Windows TaskManager], which disturb CPU and GPU calculation equally/simultaneously.

It would be better that only the CPU tasks would be disturbed, but BOINC and/or the OS (Windows) isn't smart enough.

AMD Phenom II X4 940 BE @ 4 x 3.0 GHz with 4x OCed GTX260-216.


It depend which kind of CPU and GPU/s you have.
Maybe, if you have >= 2x GTX2xx GPU series in your PC case, one/two (or more) free CPU-Core/s.

____________
BR



>Das Deutsche Cafe. The German Cafe.<

Profile Careface
Send message
Joined: 6 Jun 03
Posts: 115
Credit: 11,626,751
RAC: 0
New Zealand
Message 941536 - Posted: 20 Oct 2009, 10:10:31 UTC - in response to Message 941470.



If I would crunch also on the CPU, the GPU performance would go massively down.
Only on (whole) CPU, maybe ~ 4,000 RAC.
Only on GPUs, last stable RAC as 24/7 ~ 52,000. ~ 13,000/GPU.


If you don't mind my asking, what are your cards OC'd to? On my 3000+ I leave the CPU idle to feed the GPU (GTX216), and at 700/1480/1025 I'm getting around 9,500 RAC.

Cheers :)

Profile Chirag Patel
Volunteer tester
Avatar
Send message
Joined: 13 Sep 05
Posts: 48
Credit: 7,894,510
RAC: 4,297
India
Message 941543 - Posted: 20 Oct 2009, 12:11:26 UTC - in response to Message 941470.

AMD Phenom II X4 940 BE @ 4 x 3.0 GHz with 4x OCed GTX260-216.


It depend which kind of CPU and GPU/s you have.
Maybe, if you have >= 2x GTX2xx GPU series in your PC case, one/two (or more) free CPU-Core/s.

Wow, I only have one mobile GTX 260M (basically a shrunk desktop 9800 GT) with a P7350 CPU. That's only 112 shaders, compared to your 864! I will see what happens when I switch off one CPU core, or both.
____________

Profile Frosted
Avatar
Send message
Joined: 11 Jul 99
Posts: 83
Credit: 3,898,641
RAC: 0
Canada
Message 941547 - Posted: 20 Oct 2009, 13:32:35 UTC - in response to Message 941543.

Chirag Patel,
I could be wrong... but you might get performance if you were to update your geforce drivers. I see you have 186.x
____________

Profile Chirag Patel
Volunteer tester
Avatar
Send message
Joined: 13 Sep 05
Posts: 48
Credit: 7,894,510
RAC: 4,297
India
Message 941558 - Posted: 20 Oct 2009, 14:44:52 UTC - in response to Message 941547.

Chirag Patel,
I could be wrong... but you might get performance if you were to update your geforce drivers. I see you have 186.x

186.82 are the latest mobile drivers available officially from nVidia. I've tried LaptopVideo2Go's custom ones based on the 191.xx desktop drivers, but I've trouble with Powermizer in the past if I used those drivers (the 260M wouldn't clock down from max clocks even when the card was idle).

And call me Chirag. ;)
____________

Claggy
Volunteer tester
Send message
Joined: 5 Jul 99
Posts: 4048
Credit: 32,693,129
RAC: 604
United Kingdom
Message 941615 - Posted: 20 Oct 2009, 23:31:16 UTC - in response to Message 941558.

Chirag Patel,
I could be wrong... but you might get performance if you were to update your geforce drivers. I see you have 186.x

186.82 are the latest mobile drivers available officially from nVidia. I've tried LaptopVideo2Go's custom ones based on the 191.xx desktop drivers, but I've trouble with Powermizer in the past if I used those drivers (the 260M wouldn't clock down from max clocks even when the card was idle).

And call me Chirag. ;)


Later Beta 190.38 Notebook drivers are available in the 2nd post in this thread:

CUDA Toolkit and SDK 2.3 released

Claggy

Profile [seti.international] Dirk Sadowski
Volunteer tester
Avatar
Send message
Joined: 6 Apr 07
Posts: 7022
Credit: 59,232,833
RAC: 20,641
Germany
Message 941701 - Posted: 21 Oct 2009, 8:14:12 UTC - in response to Message 941536.

If you don't mind my asking, what are your cards OC'd to? On my 3000+ I leave the CPU idle to feed the GPU (GTX216), and at 700/1480/1025 I'm getting around 9,500 RAC.

Cheers :)


My manufacturer OCed GTX260-216 run at manufacturer OCed: 675/1458/1152 [GPU/shader/RAM].
[http://www.evga.com/products/moreInfo.asp?pn=896-P3-1258-AR&family=Geforce%20200%20Series%20Family]
[http://www.evga.com/products/pdf/896-P3-1258.pdf]

So I guess faster RAM speed would speed up the calculation.

Also, you have your GTX260-216 in a PCIe 1.0 x16 or PCIe 2.0 x8 slot?

Also faster system RAM would speed up the GPU calculation.
I changed DDR2 800/5-5-5-18 to 1066/5-5-5-15 and got ~ 5 % speed up.

____________
BR



>Das Deutsche Cafe. The German Cafe.<

Profile [seti.international] Dirk Sadowski
Volunteer tester
Avatar
Send message
Joined: 6 Apr 07
Posts: 7022
Credit: 59,232,833
RAC: 20,641
Germany
Message 941702 - Posted: 21 Oct 2009, 8:20:58 UTC


Ahh.. BTW..

You have nVIDIA_driver_190.38+ and CUDA_V2.3 installed?
This will give you ~ 30 % speed up.
[http://lunatics.kwsn.net/index.php?module=Downloads;sa=dlview;id=208]

This would explain your ~ 9,500 and my ~ 13,000 RAC.

;-)

Profile [seti.international] Dirk Sadowski
Volunteer tester
Avatar
Send message
Joined: 6 Apr 07
Posts: 7022
Credit: 59,232,833
RAC: 20,641
Germany
Message 941705 - Posted: 21 Oct 2009, 8:33:37 UTC - in response to Message 941543.

Wow, I only have one mobile GTX 260M (basically a shrunk desktop 9800 GT) with a P7350 CPU. That's only 112 shaders, compared to your 864! I will see what happens when I switch off one CPU core, or both.


But - if it's a mobile GPU in a laptop/notebook - you must look to the GPU temps.

Like I said, because of boinc.exe and System activity/peaks, CPU and GPU tasks are disturbed.

If you work on your laptop, also if you have a big WU cache, this will disturb the GPU calculation. (if the CPU isn't idle)

Because, nearly all progs have 'normal' priority also boinc.exe . CPU tasks have 'low', GPU tasks have 'lower than normal'. If a prog with 'normal' priority have activity, all under 'normal' will be disturbed. So CPU and GPU tasks. BOINC and/or the OS (Windows) isn't smart enough to disturb in the priority hierarchy of the tasks. That first only CPU tasks and if all CPU tasks stopped then GPU tasks.

--------------------------------------------

It's not possible to make BOINC more intelligent?

____________
BR



>Das Deutsche Cafe. The German Cafe.<

Profile Careface
Send message
Joined: 6 Jun 03
Posts: 115
Credit: 11,626,751
RAC: 0
New Zealand
Message 941714 - Posted: 21 Oct 2009, 10:27:10 UTC - in response to Message 941702.
Last modified: 21 Oct 2009, 10:34:22 UTC

If you don't mind my asking, what are your cards OC'd to? On my 3000+ I leave the CPU idle to feed the GPU (GTX216), and at 700/1480/1025 I'm getting around 9,500 RAC.

Cheers :)


My manufacturer OCed GTX260-216 run at manufacturer OCed: 675/1458/1152 [GPU/shader/RAM].
[http://www.evga.com/products/moreInfo.asp?pn=896-P3-1258-AR&family=Geforce%20200%20Series%20Family]
[http://www.evga.com/products/pdf/896-P3-1258.pdf]

So I guess faster RAM speed would speed up the calculation.

Also, you have your GTX260-216 in a PCIe 1.0 x16 or PCIe 2.0 x8 slot?

Also faster system RAM would speed up the GPU calculation.
I changed DDR2 800/5-5-5-18 to 1066/5-5-5-15 and got ~ 5 % speed up.


Ahh, the mem clock may account for some RAC, but I wouldn't think 3000!

EDIT: Just upped the clocks to 700/1485/1040, will see what happens..

Yeah, my card is in a PCI-e v1 x16 slot.

I'm only using DDR400 @ stock (using 11/9 divider, as the next one up is very buggy on my board) @ 3-3-3-7-1T, so again if I had your setup, I could probably get another 5% or so..

Ahh.. BTW..

You have nVIDIA_driver_190.38+ and CUDA_V2.3 installed?
This will give you ~ 30 % speed up.
[http://lunatics.kwsn.net/index.php?module=Downloads;sa=dlview;id=208]

This would explain your ~ 9,500 and my ~ 13,000 RAC.

;-)


Check.. I was almost worried I didn't have the v2.3 dlls, but a quick cmd -> fc showed that I did..

I'm wondering if its simply the fact that my CPU/RAM are getting a bit old (going to celebrate their 5 year birthday in April!)

Thankfully I almost having everything I need to piece together my i7 920 setup :) my GA-EX58-UD4P should be arriving this week, which just leaves the cpu (i7 920), and RAM (thinking of going with 6gb OCZ Reaper 1800mhz kit)

Profile Chirag Patel
Volunteer tester
Avatar
Send message
Joined: 13 Sep 05
Posts: 48
Credit: 7,894,510
RAC: 4,297
India
Message 941984 - Posted: 22 Oct 2009, 17:33:17 UTC - in response to Message 941705.

Later Beta 190.38 Notebook drivers are available in the 2nd post in this thread:

CUDA Toolkit and SDK 2.3 released

Claggy

Cool, thanks. I downloaded and installed the 190.38 beta drivers yesterday, and got the CUDA V2.3 DLLs from Lunatics, and now my long WUs finish in ~25 minutes, instead of ~45 minutes.

I am a bit concerned though, how can just a CUDA version upgrade cause such a massive increase in efficiency? I hope the Lunatics team are not taking any shortcuts in SETI@Home's algorithms just to get the highest RAC.

But - if it's a mobile GPU in a laptop/notebook - you must look to the GPU temps.

Like I said, because of boinc.exe and System activity/peaks, CPU and GPU tasks are disturbed.

If you work on your laptop, also if you have a big WU cache, this will disturb the GPU calculation. (if the CPU isn't idle)

Yes, I have to watch the temperatures sometimes. I've set the system to shut down if the GPU temp goes above 88C. The max is around 85C with the GPU at stock speeds (550/1375/950), and with the driver and CUDA DLL upgrade, the average temp is now 82C.

Also, there is absolutely no extra disk or CPU activity when I leave the laptop alone, except the occasional indexing. I don't have antivirus software installed (yes, I know the risks), and Windows 7 is very good about keeping CPU activity minimal.
____________

msattler
Volunteer tester
Avatar
Send message
Joined: 9 Jul 00
Posts: 38316
Credit: 558,976,562
RAC: 633,747
United States
Message 941985 - Posted: 22 Oct 2009, 17:38:29 UTC - in response to Message 941984.

Later Beta 190.38 Notebook drivers are available in the 2nd post in this thread:

CUDA Toolkit and SDK 2.3 released

Claggy

Cool, thanks. I downloaded and installed the 190.38 beta drivers yesterday, and got the CUDA V2.3 DLLs from Lunatics, and now my long WUs finish in ~25 minutes, instead of ~45 minutes.

I am a bit concerned though, how can just a CUDA version upgrade cause such a massive increase in efficiency? I hope the Lunatics team are not taking any shortcuts in SETI@Home's algorithms just to get the highest RAC.


Not to worry.....
The Lunatics apps are tested and validate correctly or they would not release them.

It's the upgrade to the 2.3 dll's that is giving you the large increase in performance. The same is true for desktop GPUs.

____________
*********************************************
Embrace your inner kitty...ya know ya wanna!

I have met a few friends in my life.
Most were cats.

1 · 2 · Next

Message boards : Number crunching : (x.xx CPUs, 1 CUDA)

Copyright © 2014 University of California