I can't seem to get GPUGRID to follow my BOINC Local Preferences. I've got the Minimum Work Buffer set to .02 days, and Max Additional Work Buffer set to .15 days, and I still get seven WU's for my 4 GPU core machines, even when there is 5-9 hours left on the 4 wu's that are running. The resource share is only 62.14%. and my app_config is set to run no more than 4 wu's at a time. Am I missing something? Hope someone can help. Thanks in advance, Rick
____________

I thought Intel GPUs are listed a resource type that is different than NVIDIA, and so they are naturally not included within any work fetch requests to GPUGrid.

Another possibility as to the problem might be that the server-side speed estimations aren't yet calculated. It takes 10 completed task for a given host (e.g. Computer X) on a given app's (e.g. Long-Run) given plan class (e.g. Cuda 5.5) before the server knows just how fast the host can do the tasks for that host/app/planclass combination.

Until that happens, estimations of speeds are.... often wrong.

PS: <max_concurrent> in app_config.xml currently limits the number of tasks of that app that can be running at a time, but it does NOT have any effect on how much work is requested. The best way to limit the amount of tasks queued up, is to limit your work fetch buffers.

Mine are set to:
Minimum 0.05 days
Max Additional: 0.15 days

I'd be curious what happens when you take that out, and restart BOINC. If it doesn't behave correctly, maybe we can look at it further to see if a bug existed/exists.

The estimated remaining runtime could well be incorrect (due to the recent short and long test work units all being described as 5M Gflops server side).

Any Network Usage restrictions in place?

I cant see your app_config or settings.
So, is your i7-4770's Intel HD4600 GPU being seen by Boinc (it would say in the first page of your Boinc Log file)?

If so then you can exclude the intel_GPU from being 'used' by GPUGrid (you maybe already did this).

If the intel_GPU is being seen by Boinc and you have not excluded it from the GPUGrid project, your buffer settings will mean you cache up to 4h of work, for 5 GPU's.

This might help you work out what's going on,
http://www.gpugrid.net/host_app_versions.php?hostid=157727

You can see what your different app performances are listed as for that system.
Should be going by 'Average turnaround time 0.66 days'

Right skgiven, you have a resource "intel_gpu" that GPUGrid doesn't explicitly say that it doesn't support (in sched_reply_www.gpugrid.net.xml), and so BOINC asks for work for it.

When it doesn't get any work for it, it should create an incrementing resource backoff, up to a 1-day backoff. You can see that backoff in both the Project properties, as well as the work_fetch_debug log statements (if you have that flag set in cc_config.xml)

For CPU, GPUGrid explicitly says "we don't do that resource" (<no_cpu_apps>1</no_cpu_apps> in sched_reply_www.gpugrid.net.xml), and so BOINC never asks GPUGrid for CPU work. That's why the project properties can say "Project has no apps for CPU". But GPUGrid doesn't yet do that for the intel_gpu resource type.

You can keep the GPU Exclusion in place if you'd like.
But I believe BOINC is doing its job of
- periodically asking GPUGrid for work for intel_gpu (since the project may start having tasks for it, since it doesn't explicitly say that it doesn't have apps for it!)
- creating an incrementing resource backoff when it gets no work (since the project may be temporarily out of work for that resource type).

If GPUGrid wanted to save a bit of RPC bandwidth for its users that have intel_gpu resources, I think they could add a line in the sched_reply_www.gpugrid.net.xml file to do so.

Does that make sense hopefully?

PS: You can fix your estimates by closing BOINC, opening client_state.xml, finding the GPUGrid duration correction factor, and setting it to something reasonable like 1.6

Just after deleting my exclude intel_gpu, and immediately got,
GPUGRID - Requesting new tasks for intel_gpu

- Obviously I wouldn’t get Intel work, but I thought it might change the scheduler causing it to cache up more work, some wonky way. Unfortunately I can't test for this because the 2 WU's I have running are 3h and 6h into a run and Boinc thinks they will take another 60h and 70h (due to the estimates).

Rick, you can download the Intel drivers and use it at Einstein.

I think skgiven would have to answer that one.
I'd wager, though, that you go to the Intel website, and look for graphics drivers there.

Rick, see the following thread in Einstein,
http://einstein.phys.uwm.edu/forum_thread.php?id=10215&nowrap=true#125627

Unfortunately, I guess you can't run the Intel HD processor if you already have two dual processor video cards in your box. It cost me $60 and an 80 mile trip to the computer store to find that out. But thanks for trying to help. Regards, Rick

Don't see why that would be the situation. Did they mention why?
Have you tried to use the IHD4600 to support your primary monitor?
While I don't have two dual GPU's in any system, I do have 3 discrete GPU's in one system and I can use my iHD4000 GPU to run Einstein.

Unfortunately, I guess you can't run the Intel HD processor if you already have two dual processor video cards in your box. It cost me $60 and an 80 mile trip to the computer store to find that out. But thanks for trying to help. Regards, Rick

Don't see why that would be the situation. Did they mention why, or just charge you $60 and say it can't be done?
Have you tried to use the IHD4600 to support your primary monitor?
While I don't have two dual GPU's in any system, I do have 3 discrete GPU's in one system and I can use my iHD4000 GPU to run Einstein.

I believe you need a load on the intel VGA (or DVI) port. One way is to plug a monitor in. Another way would be to use a dummy vga plug. The way most bios's are set up is if they don't detect a device on the iGPU and you have another graphics card they simply disable the iGPU

Also if you get the intel GPU driver through windows update its not usable for crunching. You need to get the 'full' package from the intel download centre.

Probably a bit late but maybe it will help someone else.

i have 4 machines running with no monitors and no dummy plugs, 3 of the machine have 2 nvidia each and 1 has 1 amd card. they all work fine. i think only the integrated gpu (intel) needs dummy plug.
____________
XtremeSystems.org - #1 Team in GPUGrid

I am happy to report successful use of two iCPU's (both 4770's). For me, it required the bios setting for the iGPU to be set to, "Always On", my monitor attached to my GTX690's, AND a VGA Dummy Plug attached to the iGPU. They have been running for 5 hrs now without a problem. So far only SETI Beta wu's have run, but hope to get Einstein wu's soon. I'll report when I get some successfully run. Getting my 3770 iGPU working using the same techniques was a dismal failure. Will keep on trying though. BTW, I'm still having the problem of getting too many GPUGRIP wu's at a time (usually 6 wu's for my 4 cores). hope someone can help me with that too. Thanks to all for the help. Regards, Rick
____________

I think, my default, you'll get "up to 2 GPUGrid tasks per GPU".

The best approach to limit BOINC work fetch from fetching too many tasks, is to:
- Make sure you are using the latest supported version of BOINC (7.0.64 currently I believe)
- Limit the buffer settings (My settings are: 0.05 days minimum buffer, 0.15 days max additional buffer)
- Use settings within cc_config.xml (in your data directory) to ensure you specifically exclude any GPUs that you do not want to do GPUGrid work. See http://boinc.berkeley.edu/wiki/Client_configuration. Regarding GPUGrid, this really should only be needed if you have any nVidia GPUs that you do not want to do GPUGrid work on.

For reference, here's my cc_config.xml file, which shows how I have excluded GPUGrid from GPU #2 (in addition to several other GPU exclusions across the 3 GPUs). Note that I have some sections commented out using xml comment blocks.

Thanks for sharing your setup Jacob. I too use a cc_config to try and exclude my iGPU from the calculations BOINC makes to get GPUGRID work. My buffer settings are similar to yours: .02 minimum days, .23 additional days. I do use project specific app_config.xml's to allow multiple projects to run simaltainiously on my GTX690's. Other than those, I let BOINC get work for all my projects and allow BOINC access to all my computer capability. Still wish I'd only get 1 GPUGRID wu per CPU, until I get to the .02 minimum shown in my preferences, and when it gets work, only get 1 WU (because 1 long wu is longer than .25 days work) also shown in my preferences. Oh well, I'll keep trying, but hate abortinh GPUGRID wu's because they've been sitting in my que unable to run because I got more than 4 wu's. Thanks again, Regards, Rick
____________

If you want me to do some more research into it, I can. What I'd need you to do is to turn on work_fetch_debug in the cc_config log_flags, then capture a segment where you believe BOINC fetched work erroneously.

I'm well-versed in reading the BOINC work fetch log files.

I figured out a reason for overfetch in my BOINC 7.1.18:

I have GPU-Grid set up as backup project, with POEM being primary and providing sporadic work supply. The cache setting is relatively low (0.05 + 0.35 days) and long runs take ~11 hours (0.46 days) on my GPU.

Yet, when POEM ran out BOINC got 2 GPU-Grids. I increased the time to switch between apps to 15h to make it finish the 1st WU which already started, but the other one may sit there for days without starting when POEM supplied work again.

So there I sat thinking why the **** does BOINC fetch 22h worth of work when it's supposed to get at most 1 for a backup project anyway? Why not wait until the 1st WU is almost finished, or at least when the expected remaining runtime falls below the cache setting? This made me switch to short tasks, where the entire problem is just not as large.

Today I think I figured out what's happening: I switched to long-runs again and got a beta WU, where this didn't happen (only fetched one, as it should). The difference on my side between beta and long run? The former is set to 1 GPU, whereas the latter is set to 0.51 to allow up to 3 POEMs along the GPU-Grid task, increasing GPU utilization from ~85% to 98%.

Apparently BOINC sees a nVidia not fully utilized when the backup project kicks in (correct, it's only 0.51), but does not yet factor in that fetching another WU of 0.51 won't help in this case. On the other hand this behaviour is not too bad, since with correct backup-project-behaviour I'd have to wait for the download of the new WU when the current one finishes. And I'd rather miss some early-return bonus than risk an idle GPU ;)

Long story short: do you have less than 1 GPU set per GPU-Grid task?

MrS