Advanced search

Message boards : Graphics cards (GPUs) : Low GPU Use

Author Message
BOINC
Send message
Joined: 26 Jan 13
Posts: 3
Credit: 7,204,825
RAC: 0
Level
Ser
Scientific publications
watwat
Message 28488 - Posted: 13 Feb 2013 | 14:05:25 UTC

So I am newish to grid computing and even newer to GPUGrid. I have a machine dedicated 24/7 for this purpose. So my question is...when I first began GPUGrid my GTX470 ran at 100% GPU load (It's liquid cooled so temp isn't a factor) but recently it only runs at 25% to 65%. I tried setting my preferences to run only the Long Run projects but that didn't help either. What am I missing here?

Core i7 2600K
8Gigs 1333MHz Ram
EVGA GTX 470
Win 7x64

My sincerest apologies if this is posted in the wrong area.
____________

Profile dskagcommunity
Avatar
Send message
Joined: 28 Apr 11
Posts: 456
Credit: 817,865,789
RAC: 0
Level
Glu
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 28491 - Posted: 13 Feb 2013 | 14:41:18 UTC

Do you hardly use your CPU parallel to gpugrid?
____________
DSKAG Austria Research Team: http://www.research.dskag.at



ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2705
Credit: 1,311,122,549
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 28501 - Posted: 13 Feb 2013 | 22:47:51 UTC

That's probably due to your GPU crunching the current beta tasks. Others are seeing similar behaviour with these. The beta is surely helpful for the project, but if you want to opt out go to your account, GPU-Grid settings and uncheck

"run test applications" (3rd from top)
"ACEMD beta:"

Not sure if both are neccessary.. but this should do it.

MrS
____________
Scanning for our furry friends since Jan 2002

BOINC
Send message
Joined: 26 Jan 13
Posts: 3
Credit: 7,204,825
RAC: 0
Level
Ser
Scientific publications
watwat
Message 28518 - Posted: 14 Feb 2013 | 15:41:38 UTC - in response to Message 28501.

Funny...thing...I didn't change anything in my settings after I made my original post..but today, automagiically I got a task that utilizes the GPU at 100%.

I guess I was being selfish and was only thinking about my numbers...I am acutally quite pleased to be using my PC tinkering projects toward something useful. I expect to bring another two systems online in the next few weeks. Thank you for the replies.
____________

ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2705
Credit: 1,311,122,549
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 28523 - Posted: 14 Feb 2013 | 20:37:18 UTC - in response to Message 28518.

The current (good) WU is probably a non-beta one. And you were right to question things when you saw much lower performance than usual - could have been that something needed fixing.

MrS
____________
Scanning for our furry friends since Jan 2002

Profile [AF>HFR>RR] alipse
Send message
Joined: 28 Oct 08
Posts: 4
Credit: 20,372,566
RAC: 0
Level
Pro
Scientific publications
watwatwatwatwatwatwatwat
Message 29007 - Posted: 6 Mar 2013 | 14:41:04 UTC
Last modified: 6 Mar 2013 | 14:41:37 UTC

Hi,

I've kinda the same problem, and i don't know how i can to
Let me explain :


I have a 670GTX and my CPU is I7 920.

When i'm loooking to GPU load, i never see more than 80% of utilization when a cpu project working in the same time.
When i stop the spu project, the GPU load going around 87%

Can some one tell me why the GPU load is never at 100%

And how i dedicated more CPU power to the GPU ? (Instead of the 0.756 cpu y defautl)?

I see that app_config can be create but what should i write in it ?

<app_config>
<app>
<name>acemdlong</name>
<max_concurrent>1</max_concurrent>
<gpu_versions>
<gpu_usage>1</gpu_usage>
<cpu_usage>1</cpu_usage>
</gpu_versions>
</app>
</app_config>

Is that correct ?



Many thanks to the community :)

ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2705
Credit: 1,311,122,549
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 29015 - Posted: 6 Mar 2013 | 19:30:08 UTC - in response to Message 29007.

You won't need an app_config for this, just set your BOINC to use at most 99% of CPUs. This will let you run 7 CPU tasks + 1 GPUGrid, which should be sufficient to fuel your GPU.

Achieving 100% GPU load is rather difficult. Previous work here achieved ~96% for me, but this varies with WU batches. WUs from the long runs queue should generally fare better than short ones, I suppose.

MrS
____________
Scanning for our furry friends since Jan 2002

Profile [AF>HFR>RR] alipse
Send message
Joined: 28 Oct 08
Posts: 4
Credit: 20,372,566
RAC: 0
Level
Pro
Scientific publications
watwatwatwatwatwatwatwat
Message 29076 - Posted: 8 Mar 2013 | 9:28:21 UTC

Thanks for ur answer, but i've already tried this tip, and if i let one core free with this option, this core won't be used by any BOINC application, and the GPU load will not changed at all.

ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2705
Credit: 1,311,122,549
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 29087 - Posted: 8 Mar 2013 | 21:40:21 UTC - in response to Message 29076.

Wouldn't this give you 7 CPU tasks and 1 GPUGrid tasks, which should now use ~50% of one logical core? So in sum you'd be using 7.5 of 8 cores. Well.. and if the crunching speed is still too much lower than without a CPU project I'm afraid you'll have to lower the number of cores used for CPU projects further.

MrS
____________
Scanning for our furry friends since Jan 2002

Profile [AF>HFR>RR] alipse
Send message
Joined: 28 Oct 08
Posts: 4
Credit: 20,372,566
RAC: 0
Level
Pro
Scientific publications
watwatwatwatwatwatwatwat
Message 29095 - Posted: 9 Mar 2013 | 14:59:00 UTC
Last modified: 9 Mar 2013 | 15:22:23 UTC

I don't know why, but when i try to set one core free on project's preference, that don't work in the manager. Still 8 cores are used and then 80% of gpu load. :/

Jim1348
Send message
Joined: 28 Jul 12
Posts: 819
Credit: 1,591,285,971
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 29097 - Posted: 9 Mar 2013 | 15:24:28 UTC - in response to Message 29007.

Can some one tell me why the GPU load is never at 100%

And how i dedicated more CPU power to the GPU ? (Instead of the 0.756 cpu y defautl)?

None of my cards (GTX 560, 650 Ti and two 660s) ever run at 100% either; that seems to be just the way the program works. And if you use your card as a display card, that is a good idea; it prevents mouse and keyboard lag (in fact, they might do it for that purpose, but I don't know).

However, a simple way to get more CPU support is to turn off hyperthreading. You boot into the BIOS for that, and there should be an option in the CPU section. Some PCs don't provide that option, but they are usually laptops, which is not your case!

By the way, there is something strange about how BOINC (7.0.54 x64) manages cores; I have to set the CPU pecentage to 75%, or else I get 5 projects running on 4 cores (with HT off), and the 5th one runs very slowly. But that may be due to the fact that I am running two GTX 660s on that PC, and that might affect how the CPU cores are counted.

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 29101 - Posted: 9 Mar 2013 | 21:50:30 UTC - in response to Message 29097.

When i'm loooking to GPU load, i never see more than 80% of utilization when a cpu project working in the same time.
When i stop the spu project, the GPU load going around 87%
Can some one tell me why the GPU load is never at 100%

That means that the maximum GPU Utilization you can get running that task type on your setup (hardware + Operating system) is 87%.
It's been a long standing fact that WinXP and Linux performances are higher (11%+) than Vista and W7; GPU Utilization from WinXP and Linux is higher than W7, Vista, W8, and even 2008 and 2013 servers. AFAIK this has not changed with recent Beta's and new apps.
It's likely that you would need to 'free' 2 cores on an i7-920 to optimize for your GTX 670.
____________
FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2705
Credit: 1,311,122,549
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 29104 - Posted: 10 Mar 2013 | 13:16:35 UTC - in response to Message 29095.

I don't know why, but when i try to set one core free on project's preference, that don't work in the manager.

I wouldn't use the gobal settings for this, since you'll likely want this to be different on all machines. And if you already set something locally (some time ago, maybe) this will override the global setting, so go for the local one directly.

However, a simple way to get more CPU support is to turn off hyperthreading.

I advice against this: it's better to leave to HT on and restrict your BOINc to use 50% of all CPUs. This way Win will still scheduler all 4 BOINC tasks on the 4 physical cores. But if you want to use your PC for anything else while crunching you'll get better performance from the remaining 4 free (logical) cores. This is an oversimplification, but I think the idea is clear: you get better automatic load balancing.

MrS
____________
Scanning for our furry friends since Jan 2002

Jim1348
Send message
Joined: 28 Jul 12
Posts: 819
Credit: 1,591,285,971
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 29106 - Posted: 10 Mar 2013 | 14:39:43 UTC - in response to Message 29104.
Last modified: 10 Mar 2013 | 14:57:08 UTC

However, a simple way to get more CPU support is to turn off hyperthreading.

I advice against this: it's better to leave to HT on and restrict your BOINc to use 50% of all CPUs. This way Win will still scheduler all 4 BOINC tasks on the 4 physical cores. But if you want to use your PC for anything else while crunching you'll get better performance from the remaining 4 free (logical) cores. This is an oversimplification, but I think the idea is clear: you get better automatic load balancing.

MrS

True, but I am in the opposite situation with that machine, which runs only BOINC/World Community Grid and GPUGrid projects, and all the cores are dedicated to them anyway. Then, it is just a question of efficiency. Some of the WCG projects work well with hyperthreading, but the one I am using at the moment (CEP2) doesn't seem to gain from it, and in fact I think it is better with HT off.

However, 8 virtual cores does give you finer granularity over how you allocate the cores, and having only 4 real cores means that I have to devote half of them to supporting my two GTX 660s for GPUGrid (they each take one core, whether real or virtual). The GPUs gain only a little in speed with the extra CPU support, but overall, I like it better, since the GPUs are high-efficiency computing devices and a small percentage increase in their output is better than devoting a couples of virtual cores to other projects in my opinion.

There is also a question of stability, which may or may not apply here. On some of the GPU projects that use OpenCL, sometimes the work units crash apparently due to lack of CPU support. Some people have reported that devoting a whole CPU core to their GPU helps. Whether that applies to CUDA I don't know, but this projects uses a lot of CPU power to support the Kepler cards, so I think it is possible

ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2705
Credit: 1,311,122,549
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 29118 - Posted: 11 Mar 2013 | 21:24:22 UTC - in response to Message 29106.

Some of the WCG projects work well with hyperthreading, but the one I am using at the moment (CEP2) doesn't seem to gain from it, and in fact I think it is better with HT off.

True, some software doesn't benefit at all or even takes a performance hit (the latter has become slightly better with each new CPU generation and Haswell might improve HT scaling significantly due to the much "fatter" core). However, I don't agree on your 2nd sentence: I'm pretty sure performance will be about equal on a modern OS (it has to know about HT). We'd have to measure it.. but I think it's not that important.

And I agree that it's a good idea to sacrifice some CPU throughput to feed a GPU properly. In fact, my i7 3770K @ 4.1 GHz is usually completely dedicated to feeding my GTX660Ti running 8 POEM WUs :D
(BTW: I don't think stability is an issue of CPU support for POEM OpenCL. Performance surely is, on the other hand.)

Now the interesting question is: since the GPUGrid threads feeding the GPU need exclusive access to a logical core to service the GPU exactly when it needs support, but don't do any number crunching themselves, would they be slowed down by another BOINC thread running on the other logical core? Say you've got a Quad with HT and run 4 WCG + 2 GPUGrid instead of 2 WCG + 2 GPUGrid. Each WCG should be a bit slower in config 1, but overall throughput will almost double. And I suppose GPUGrid performance would hardly suffer.. but this would have to be measured properly.

MrS
____________
Scanning for our furry friends since Jan 2002

Jim1348
Send message
Joined: 28 Jul 12
Posts: 819
Credit: 1,591,285,971
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 29119 - Posted: 11 Mar 2013 | 22:21:17 UTC - in response to Message 29118.

Now the interesting question is: since the GPUGrid threads feeding the GPU need exclusive access to a logical core to service the GPU exactly when it needs support, but don't do any number crunching themselves, would they be slowed down by another BOINC thread running on the other logical core? Say you've got a Quad with HT and run 4 WCG + 2 GPUGrid instead of 2 WCG + 2 GPUGrid. Each WCG should be a bit slower in config 1, but overall throughput will almost double. And I suppose GPUGrid performance would hardly suffer.. but this would have to be measured properly.

MrS

It depends a lot on the I/O requirements of that particular WCG project. CEP2 takes a LOT of I/O. When I ran it on my old Q9450 quad-core, I could see the reduction in CPU efficiency (via BoincTasks) as I increased the number of tasks running simultaneously from 1 to 4. But with my Ivy Bridge CPUs (i5-3550 and i7-3770) it is much less evident. And the other WCG projects don't have much affect on the other cores.

Adding an Nvidia Fermi card (e.g., GTX 560) to that takes away a little, maybe 1% or so from each CPU core, but the Fermi cards do not demand exclusive access to a CPU core for some reason. My guess at the moment, though I have not checked it out carefully is that a Kepler card would not take anything away from the other WCG projects, since it already has exclusive access to its own core and has low I/O requirements.

Profile chip
Send message
Joined: 10 Feb 09
Posts: 5
Credit: 50,333,581
RAC: 0
Level
Thr
Scientific publications
watwatwatwatwatwatwat
Message 29162 - Posted: 14 Mar 2013 | 10:01:39 UTC

GTX470 @ 700MHz, i7-2600K @ 4500MHz
I3R52-NATHAN_dhfr36_3-3-32-RND1460_1_9
0.19 CPU [acemd.2865P.exe]
0.04 ISR/DPC [dxgkrnl.sys + nvlddmkm.sys]

85% - this is normal?

Jim1348
Send message
Joined: 28 Jul 12
Posts: 819
Credit: 1,591,285,971
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 29163 - Posted: 14 Mar 2013 | 10:15:26 UTC - in response to Message 29162.

85% - this is normal?

My GTX 560/650 Ti/660 tend to run at 92-93%, but I have seen them at 85% often enough that I think it is quite normal.

ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2705
Credit: 1,311,122,549
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 29179 - Posted: 16 Mar 2013 | 13:17:52 UTC - in response to Message 29119.
Last modified: 16 Mar 2013 | 13:18:02 UTC

It depends a lot on the I/O requirements of that particular WCG project. CEP2 takes a LOT of I/O. When I ran it on my old Q9450 quad-core, I could see the reduction in CPU efficiency (via BoincTasks) as I increased the number of tasks running simultaneously from 1 to 4. But with my Ivy Bridge CPUs (i5-3550 and i7-3770) it is much less evident.

That was probably main memory bandwidth and latency, which both have been tremendously improved by the first "Core i" series with their integrated memory controller. Or actually the Quads, as the dual cores back then still had the old controller, just on die.

My guess at the moment, though I have not checked it out carefully is that a Kepler card would not take anything away from the other WCG projects, since it already has exclusive access to its own core and has low I/O requirements.

Yes, that's what I expect as well :)

MrS
____________
Scanning for our furry friends since Jan 2002

Profile X1900AIW
Send message
Joined: 12 Sep 08
Posts: 74
Credit: 23,566,124
RAC: 0
Level
Pro
Scientific publications
watwatwatwatwatwatwatwatwatwat
Message 29187 - Posted: 17 Mar 2013 | 9:33:41 UTC - in response to Message 29162.

GTX470 @ 700MHz, i7-2600K @ 4500MHz

GTX 650TI (fan fixed to 40%) + i7-3770K @ 4.0 Ghz
GPU load: 94 percent.
BOINC: 0.528 CPUs + 1 GPU

Post to thread

Message boards : Graphics cards (GPUs) : Low GPU Use

//