Thanks for your effort, Jacob!

That ~3% hit in GPU performance is a number hard enough for me to use as a future guideline. I wouldn't expect this value to change significantly upon the various changes SK mentioned. And for most other CPU tasks this value should apply.. let's call them well-behaved. There may be tasks which are much more demanding (e.g. Climateprediction), but these are the exception rather than the norm.

However, there should be an important trend: the performance hit should grow, the faster the GPU becomes, as it would need CPU support more frequently (-> more interruptions) and a similar delays waste more GPU power (this could actually already be accounted for by measuring the performance hit in %).
And put the other way around: the slower a GPU, the less important the CPU support should be. So people with smaller GPUs can feel free to use their CPUs elsewhere.

And if you're up for more tests there's one thing I think is missing: a comparison between 7 CPU tasks and 8, in addition to comparing to 0 CPU tasks. You may find the same 3%, which would make the choice tougher: my GTX660Ti is good for ~324k RAC crunching long-runs, so 3% less would cost me 9.7k RAC. That's about what a fast CPU can achieve running well-paying projects.. but nowhere near the throughput of a single virtual core (whose companion is already occupied).

MrS

---

Scanning for our furry friends since Jan 2002

I will probably perform your requested test sometime soon. Right now, though, I'm testing running 2 tasks at the same time on the same GPU, and actually having some success.

When I have results ready, I think I'll create a new thread, so that posts to this thread can remain related to app_config.

The previous result was a speedup of ~5%.. let's see what you get!

MrS

---

Scanning for our furry friends since Jan 2002

The thread that contains my performance tuning can be found here:
http://www.gpugrid.net/forum_thread.php?id=3331

========================================================================
As I mentioned before:
To maximize my crunching efforts, I'm using and recommending:

BOINC settings:
% of the processors: 100%
% CPU time: 100%

app_config settings for each of the 4 GPUGrid projects:
<max_concurrent>9999</max_concurrent>
<gpu_usage>1</gpu_usage>
<cpu_usage>0.001</cpu_usage>

I'm still testing on "doubling-up" GPU tasks on a single GPU; if you wanted to do this, you'd use:
<gpu_usage>0.5</gpu_usage>

Best of luck,
Jacob

FYI:

Per Toni (a project administrator), because of application and validator problems, I was told to disable running 2-at-a-time. So I have suspended my testing/research, and have put <gpu_usage> to 1, in the app_config.xml file.

http://www.gpugrid.net/forum_thread.php?id=3332#29425

Regards,
Jacob Klein

FYI:

Per Toni (a project administrator), because of application and validator problems, I was told to disable running 2-at-a-time. So I have suspended my testing/research, and have put <gpu_usage> to 1, in the app_config.xml file.

http://www.gpugrid.net/forum_thread.php?id=3332#29425

Regards,
Jacob Klein

I have solved the problem I was having where Nathan tasks were not processing correctly on my machine.
The problem was completely unrelated to the app_config.xml file.
Details here: http://www.gpugrid.net/forum_thread.php?id=3332&nowrap=true#29894

So, we can resume app_config testing (including 2-tasks-on-1-GPU).

I still recommend using the app_config.xml file that is in this post:
http://www.gpugrid.net/forum_thread.php?id=3319&nowrap=true#29216
... and I only recommend trying 2-tasks-on-1-GPU if you are running GPUGrid on GPUs that have 2GB or more RAM; if that's the case, you might try using <gpu_usage> value of 0.5 in your app_config.xml file, to see if GPU Load increases (by using GPU-Z) and throughput increases (by looking at task Run Times).

Thanks,
Jacob

i know but i want use 87% 1st wu and 13% 2nd wu , is that possible ?

If you are asking if you can tell BOINC to only use a specified portion of a GPU for each of the tasks that are running on the GPU, then answer is no.

You can only use <gpu_usage>, to have BOINC determine how many tasks to run on a single GPU. You cannot specify that certain tasks should be run more of the time.

Regards,
Jacob

thx for the reply, very sad, what i have to do with this saved power, i dont want to slow down my ending times of a wu ?

Just thought , can i run a second instance of boinc in the same inviroment and start a second gpugrid instance ???

Or is there a way to start a command line option to run gpugrid wus ?

Thanks for asking the same thing in the other thread, so that I jsut wrote about the same as Jacob over here..

Regarding your other questions: it might be possible if you could launch the 2nd task at lower priority, then elevate it to normal when it becomes the 1st one (otherwise performance will suffer from CPU projects). I don't know how to do this, though. Except modifying BOINC yourself.

MrS

---

Scanning for our furry friends since Jan 2002

There is a software on this site : BoincTasks

Where u can adjust the priority of the tasks, allready installed yet, but still trying.

i thought wrong, there is no adjustment.

I will try Windows Task Manager...no gpu usage change by changing the priority with windows. clueless :(

Yeah, but you'd have to assign different priorities to tasks of the same name, and switch priority while the task is running. You could probably ask the developer to include such functionality, I don't suppose it's already there.

MrS

---

Scanning for our furry friends since Jan 2002

As far as I have seen, a single task will only use "so much" of the GPU's "Usage" or "Load".

It's normal for GPU "Usage" or "Load" to vary, per app, and per project. I've seen some cases where the task makes it 99% (some SETI tasks), some where it's around 80-90% (Nathans here at GPUGrid), and some where it's quite low like 40% (POEM tasks).


To increase a GPU's "Usage" or "Load", so far as I know, your options are:
- Free up CPU resources, so the GPU is never waiting on the CPU to help it get a kernel of work started/ran.
- Run more tasks at once on the GPU, if possible

I thought a GPU used some of it's own processing power for house keeping purposes, I've noticed when a wu breaks, the GPU will shoot up to 100% while the memory controller goes to 0%. So the more the data is being pumped through the GDDR memory, the more the GPU usage will lower slightly, amongst other things.

ok, i will give the gpu a full cpu and see what happens next.

No effect, still the same gpu usage...
we must be waiting for a solution...

so long...

GPUGRID WU's usually use the GPU at 98 to 99% when running on Linux, XP or a 2003 server.
Utilization usually only drops on WDDM operating system (Vista, W7, W8, 2008 and 2013 Windows servers). On 2008 server it's slightly less (~5% instead of 11% with W7 last time I measured it) but different system architectures have different bottlenecks.

There are lots of other little things you can do (setting tweaks) to reduce the system overhead/maximize GPU performance; disable Aero, set your systems performance settings to adjust for best performance, disable unnecessary startup apps, use faster system memory, a SSD (though it might not last several years), a RAM drive or use a secondary drive for Boinc, use faster RAM (2133/2400 rather than 1333/1600), OC the CPU slightly, OC the GPU (especially the GDDR, as this should increase GPU usage), make sure it's using PCIE X16 as opposed to X8 or X4, use an iGPU for display, set the GPU to prefer maximum performance, increase the fan speed so that the GPU stays cooler, boosts well and doesn't downclock, use a better PSU (which apparently aids the Boost and keeps the system cooler), make sure your GPU is running at PCIE3 rather than dropping down to PCIE2 because your other GPU is PCIE2 only capable, stop using more than one CPU core/thread if your GPU and CPU time are not the same (or almost) - 33,047.69 vs 31,835.09 might represent 3.8% loss due to the CPU usage. Also, if you don't OC your CPU it's frequency isn't the same when one core is in use of 4 cores (3.5 vs 3.9GHz is an 11% CPU difference)...

---

FAQ's

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

Hi, guys.

Having some issues getting multiple tasks to run per GPU.

I'd like (if possible) to have 2 tasks per GPU running of any and all of the GPUGrid tasks available. I'm not sure how to set this up regarding what the <name> value(s) should be currently in the app_config file.

If only one task/work unit type is available per value, I would prefer the long runs.

Any insight would be appreciated.

Thanks!

...

Edit: If there's some way to set this up globally in BOINC for all projects instead, that would be ideal. I already have BOINC set up to run 2 GPU tasks (presently 1 per GPU) simultaneously instead of the default 1.

To clarify: I'm trying to get a total of 4 GPU tasks to run simultaneously.

---

My BOINC Cruncher, Minecraft Multiserver, Mobile Device Mainframe, and Home Entertainment System/Workstation: http://www.overclock.net/lists/display/view/id/4678036#

The <name> values can be found within your client_state.xml file, assuming you had previously downloaded a task for the application. Make sure that you do not touch/edit/save that file... just look at it, don't change it :)

Then the BOINC configuration for setting up the app_config can be found here:
http://boinc.berkeley.edu/wiki/Client_configuration

Finally, it sounds like you might be looking for the following (pasted below), which could be placed in your app_config.xml file within the gpugrid project folder. The gpu_usage of 0.5 indicates 2 tasks per GPU, and the cpu_usage of 0.667 means reserve 1 core when 2 tasks are running, and 2 cores when 3 tasks are running.

Good luck!
Jacob

<app_config>

   <!-- Short runs (2-3 hours on fastest card) -->
   <app>
      <name>acemdshort</name>
      <max_concurrent>0</max_concurrent>
      <gpu_versions>
          <gpu_usage>0.5</gpu_usage>
          <cpu_usage>0.667</cpu_usage>
      </gpu_versions>
    </app>

   <!-- Long runs (8-12 hours on fastest card) -->
   <app>
      <name>acemdlong</name>
      <max_concurrent>0</max_concurrent>
      <gpu_versions>
          <gpu_usage>0.5</gpu_usage>
          <cpu_usage>0.667</cpu_usage>
      </gpu_versions>
    </app>

   <!-- ACEMD beta version -->
   <app>
      <name>acemdbeta</name>
      <max_concurrent>0</max_concurrent>
      <gpu_versions>
          <gpu_usage>0.5</gpu_usage>
          <cpu_usage>0.667</cpu_usage>
      </gpu_versions>
    </app>

</app_config>

Thank you, sir! And thanks for the prompt reply. It's much appreciated.

I modified the other values you listed a little for my needs.

So far this seems to be working as desired. More tasks are being pulled in (downloaded) now, and the two GPU tasks I had running are now running on one GPU.

Rough estimate so far is that my total VRAM and VRAM per card will only be about 50% loaded running the 4 tasks simultaneously, which leaves more than enough head room even for gaming and the like. I even have the two cards in SLI configuration so their memory usage is mirrored.

Very awesome, and should give a significant boost to the total work per time I'm able to crunch on this computer.

Thanks again.

---

My BOINC Cruncher, Minecraft Multiserver, Mobile Device Mainframe, and Home Entertainment System/Workstation: http://www.overclock.net/lists/display/view/id/4678036#

Setup info:

I have two EVGA Titan Black Superclocked cards (6GB VRAM each). They are set to SLI in the Nvidia control panel for game rendering optimization. This causes their VRAM to be mirrored between the two cards so they can work on the same data at the same time.

In the Nvidia control panel I also have prefer maximum performance set to basically help keep the GPU clock rates up so that they're able to do more work per time.

I also have the C:\ProgramData\BOINC\cc_config.xml <use_all_gpus>1</use_all_gpus> value set to use each of my GPUs instead of just one.

OK, upon further testing:

With 4 GPUGrid tasks loaded in VRAM on each card (since the VRAM is mirrored), the VRAM is about 30% to 40% loaded on average, and with 2 tasks running per GPU, they're about 80% to 90% loaded on average (this is up about 20% from when only 1 task is being run per GPU). Additionally, the GPU clock rates seem to be much more consistently held at their higher stock boosted clock rate of 1125MHz. (I have a custom fan curve set for these cards to keep them well and cool and nowhere near throttling down.)

This should yield a significant performance increase in regard to the amount of overall work done per time.

...

I generally use 67% (8 of 12) CPU threads to crunch for BOINC, leaving 4 open for other tasks and programs. So I now simultaneously run 4 GPU tasks (GPUGrid) and 8 CPU tasks (World Community Grid).

So given that, what would be the optimal value for <cpu_usage>? I've tried 0.125, 0.25, and 0.5. I would prefer to leave enough room open on my CPU for other programs, but wouldn't mind using up to the equivalent of one additional thread as needed for the GPU tasks, so should I set this value to 0.25? But then again, would they even really use it? What is the default amount used for GPUGrid GPU tasks?

Thanks.

---

My BOINC Cruncher, Minecraft Multiserver, Mobile Device Mainframe, and Home Entertainment System/Workstation: http://www.overclock.net/lists/display/view/id/4678036#