PAOLA_3EKO_8LIGANDS

On both my GTX670's, I'm averaging 30% GPU load and the core has downclocked to 1005MHz. At that rate, it looks like it's going to take 12 to 14 hours to completion for those wu's (this is the first time I've had these work units). Anyone else having these problems?

I have 3 of these units running now and yes, they are very, very slow, around 30% utilization on windows 7, and 37% on windows xp.

Same behavior with a gtx570, 38% after 5 hours, another 6 estimated, gpu usage at ~48%

Same problem. Cuda31 285gtx 26 hours to complete at 46% gpu load and 33% cpu load

---

DSKAG Austria Research Team: http://www.research.dskag.at

For this group, here is a list of the GPU and runtime for the 22 most recently returned WUs:

GeForce GTX 680 - 8.6 hours
GeForce GTX 680 - 16.9 hours
GeForce GTX 680 - 15.8 hours
GeForce GTX 680 - 15.1 hours
GeForce GTX 680 - 14.8 hours
GeForce GTX 670 - 16.5 hours
GeForce GTX 670 - 14.0 hours
GeForce GTX 670 - 13.3 hours
GeForce GTX 590 - 15.1 hours
GeForce GTX 590 - 14.8 hours
GeForce GTX 590 - 12.8 hours
GeForce GTX 580 - 9.6 hours
GeForce GTX 580 - 15.3 hours
GeForce GTX 580 - 13.6 hours
GeForce GTX 570 - 22.3 hours
GeForce GTX 570 - 18.0 hours
GeForce GTX 570 - 15.4 hours
GeForce GTX 570 - 14.7 hours
GeForce GTX 560 - 15.8 hours
GeForce GTX 560 - 16.4 hours
GeForce GTX 470 - 17.5 hours
GeForce GTX 470 - 17.3 hours

Lots of long runtimes, and a lot of variablity. Two 470s running almost as fast as a 680? The ones in bold are what we were expecting based on benchmarks here, and show what "ideal" conditions. Any ideas as to why so many are running slow? It's obviously not good that so many fast cards/machines are running slow for some reason. Hmmm...

GTX460 768MB 62% GPU usage completed in 19.14

GTX670: 95.460% 16:32:28 hores runtime.
GPU load between: 27% and 31%, CPU load 9% on an AMD octocore. One core is recerved for GPUGRID and work, the other 7 are doing Climateprediction.net.

UP-DATE: the WU was complited after 17:22:56 hours.

Hope this helps.

Observationally on GTX670 Win7x64 the CPU usage is at a lower ratio on these PAOLA_8LIGANDS (1 GPU sec to .7 CPU sec) than the NATES (1:1).
I guessing that the GPU is starving while waiting for the CPU. I don't know how you have it coded (CPU is polling or the GPU is placing a call) but if this is configurable any chance you could switch it so we can test this for you?

So maybe we can look at general system states to make sure that the extra processing runtimes are not caused by the unexpected requirements that these computationally difficult WUs have? Anyone at their system at this time can you take a look at how much GPU and also system RAM is being used and how much is reported as free for each? How about cpu page faults?

As always, please include OS and any other pertinant system details.

I'd be happy to test anything new or even crunch them as they are for the project but please throw us crunchers a bone and up the points per WU :-)

---

Thanks - Steve

nate got me to thinking, it's taking the same amount of time on my GTX 560's and GTX 550's as it does on my GTX 670's. 1344 cuda cores vs 192 and it takes almost an identical amount of time, somethings wrong.

I run Docking@home in addition to GPUGRID on my Phenom II X6. I suspended Docking to observe how much CPU resources was being consumed by my GTX460/GTX550Ti whilst these units are running and it turned out to be 33%.

Reduced Docking down to 4 cores to see what effect it has.

460 running at 60% GPU usage and 550Ti running at 99%

Win XP

Reduced Docking down to 4 cores to see what effect it has.

460 running at 60% GPU usage and 550Ti running at 99%

Wow, you just now figured that out? Maybe I'm not understanding what you wrote, I thought everybody knew that you need to leave 1 CPU core free for every GPU you're running. The CPU core feeds the data to the GPU, it's going to be like Christmas for you now, you're crunching times should drop sharply.

---

Reduced Docking down to 4 cores to see what effect it has.

460 running at 60% GPU usage and 550Ti running at 99%

Wow, you just now figured that out? Maybe I'm not understanding what you wrote, I thought everybody knew that you need to leave 1 CPU core free for every GPU you're running. The CPU core feeds the data to the GPU, it's going to be like Christmas for you now, you're crunching times should drop sharply.

You don't need to leave 1 CPU core free per GPU, but it helps. I have chosen to run 1 free core for 2 GPU. I find it is the most efficient was to use my resources.

Docking is my main project....... 20th overall. GPUGRID is only a side project for me that compliments D@H very well, they are both in the same area of research.

Used to run 6 cores Docking at the same time as GPUGRID, wasn't bothered about how much it affected GPUGRID crunch times but when I started running 2 GFX cards on GPUGRID it started hitting D@H crunch times, so I reduced to 5 cores which didn't really affect my D@H RAC.

Whilst these CPU intensive PAOLA units are around I'll run D@H on 4 cores, then I'll probably return to 5.

Previous to these PAOLA units I was getting ~96% CPU sage overall with 5 D@H units and 2 GPUGRID WU running. Currently seeing 97% with 4 D@H + 2 PAOLA WU.

I figured you must have known, I get 0.645% CPU usage per core on my GTX 670's, using 2 cards in the same machine(do the math, that's more than one CPU core). I'm leaving 2 of my 8 cores free and the GPU usage shot up to 98% on both cards.

Before that it jumped all over the place between 35 and 65% and the wu runs were hours longer (I was a n00b, I didn't know). I run CPDN on all other cores, when their servers go down (witch is often) I run docking too. I like those little 3 hour jobs, for some reason it wont link to my BOINCstats, I must have done something wrong there too.

Edit: My GTX670 took 18.27.15 and produced 82.54MB of data

GTX560Ti took 19.24.51 and produced 82.65MB of data

---

I figured you must have known, I get 0.645% CPU usage per core on my GTX 670's, using 2 cards in the same machine(do the math, that's more than one CPU core). I'm leaving 2 of my 8 cores free and the GPU usage shot up to 98% on both cards.

Before that it jumped all over the place between 35 and 65% and the wu runs were hours longer (I was a n00b, I didn't know). I run CPDN on all other cores, when their servers go down (witch is often) I run docking too. I like those little 3 hour jobs, for some reason it wont link to my BOINCstats, I must have done something wrong there too.

Edit: My GTX670 took 18.27.15 and produced 82.54MB of data

GTX560Ti took 19.24.51 and produced 82.65MB of data

I see you run Bulldozers. On Docking the 6 core Phenom II's outperform the 8 core Bulldozers. One of the reasons I have stuck with my old skool Phenoms

My main cruncher - a Phenom II X6 1055T i run overclocked at 3.5GHz, at that speed running 5 cores on Docking is just as effective as running 6 cores at 2.8GHz.

These workunits do not use a full CPU core with Kepler GPUs, unlike any previous workunits. It's like the late swan_sync parameter wasn't set to 0. These workunits run twice as fast on my GTX 480s than on my GTX 680s.

WinXP x64, Core i7 980x and 970 (@ 4.16GHz) one free CPU thread per GPU, GTX 680s GTX 690 and GTX 480

I'm running Windows 7x64 with a single GTX 670 and an i7-3770K overclocked to 4.2GHz with hyperthreading enabled (i.e. 8 logical cores). My card is the factory-overclocked triple-fan Gigabyte 670. All factory default settings. I'm running BOINC 7.0.28. No cc_config, no swan_sync.

I've run one Paola 3EKO task so far and the result was validated yesterday. Here are the specs:
Run Time (secs):48,205.04
CPU Time (secs):30,510.69
Credit :88,075.00

I'm currently running another Paola 3EK0, and the GPU load right now is fluctuating between 44-46% with the core clock at 979.8 MHz. This is while also running 8 tasks of World Community Grid on this computer at the same time, so I haven't dedicated a core to GPUGrid (it's sharing cores with WCG).

If I suspend WCG completely in BOINC (thus leaving the *entire* CPU for GPUGrid, the GPU load rises to 53%.

If I set BOINC to use 87.5% of processors (that's equal to 7 out of 8 cores), it shuts down 1 WCG task (so only 7 remain running), but GPU load remains 45%.

If I set BOINC to use 75.0% of processors (that's equal to 6 out of 8 cores), it shuts down two WCG task (so only 6 remain running), but GPU load still remains 45%.

If I set boinc to use 50% of processors... ditto.

With WCG shut down, the GPUGrid task, in task manager (Ctrl+Alt+Delete) reports a CPU usage of 6-7%, so in my case that's only half a core. The priority is set to "BelowNormal" by default, whilst for WCG each task has a priority of "Low" when running. I don't understand it. It seems that the Paola tasks are affected only by the fact that another project is running - not by how much CPU that project takes. So you have to suspend ALL other tasks *completely* to see a slight performance increase.

Nathan tasks have 90% GPU utilisation regardless of what else is running. These Paola tasks are using only half my GPU and returning about half the points per unit time. According to my results, I get about 4 points per second for Nathan, one point per second for Noelia, and 2 points per second for the new Paola ones. I'm not particularly happy to be running at half throttle when I know I could get more points per unit time doing other work. In fact I've started considering other projects like primegrid or POEM@home to maybe increase my points per hour. I guess I'll just see how things unfold and take it from there.

does anyone has default max 20% CPU time for GPU work set on the websiteprofiles too like me until today? perhaps this value is to low for these new units? i set it to 100% and waiting now until i get a new one of this WUs and finished it.

---

DSKAG Austria Research Team: http://www.research.dskag.at

does anyone has default max 20% CPU time for GPU work set on the websiteprofiles too like me until today? perhaps this value is to low for these new units? i set it to 100% and waiting now until i get a new one of this WUs and finished it.

I feel it must be something like this, because there are some users who can compute much faster than the rest (and at speeds we were expecting). Keep us updated dskagcommunity.

If anyone else wants to play with the setting, click on your username up above, then "GPUGRID preferences". "Edit Preferences", and change "Maximum CPU % for graphics..." to 100% (or whatever you prefer).

Still, this might not be it. Wouldn't explain this, though, unless the cards are on different machines with different settings...

These workunits do not use a full CPU core with Kepler GPUs, unlike any previous workunits. It's like the late swan_sync parameter wasn't set to 0. These workunits run twice as fast on my GTX 480s than on my GTX 680s.

Let's see...

If I set BOINC to use 87.5% of processors (that's equal to 7 out of 8 cores), it shuts down 1 WCG task (so only 7 remain running), but GPU load remains 45%.

If I set BOINC to use 75.0% of processors (that's equal to 6 out of 8 cores), it shuts down two WCG task (so only 6 remain running), but GPU load still remains 45%.

If I set boinc to use 50% of processors... ditto.

I've seen people advise this action here and at other BOINC forums and it seems to me that this would never work because telling BOINC to use 6 of 8 cores or 7 of 8 cores takes them away from all projects. I would think you would want to set CPU usage at you're GPUgrid account, by taking away cores in BOINC, only the operating system or programs not connected to BOINC can utilize those cores.

I don't think the prefrences in our GPUGRID account allows for enough minipulation of the CPU to make it do what you want.

---

Here's the one thing I've been able to find in common with all my video cards, the GPU memory controller stays right around 10%, it will drop to 9% and go up to 11% but never higher or lower on the 3EKO wu's. Also, they are all using the same amount of memory (+ or - 1%) witch is 628MB, these are my cards:

2 EVGA GTX550Ti 1024MB GDDR5
3 EVGA GTX560 1 being a Ti all have 1024MB GDDR5
2 EVGA GTX670FTW 2048MB GDDR5

I know the CPU feeds data to the VRAM, right? If there is a bottle neck in the way the data is using the VRAM, that will slow down the CPU and GPU both. Everybodys cards have different GPU clock speeds, so wouldn't they down clock if there was a data bottle neck.

Could it have something to do with the way those wu's are utilizing the VRAM? I think the problem is in that area some where.

---