tips, tricks, and tidbits for GPU crunching

[roki977]

When I came home from work today, I see this. Project has no jobs available. There was no reset or instability. Is this normal, only one job done today?



Something wrong,"communucation deferred 0..."ž

This is frustrating. I can not do anything to start the GPU Grid again.

He went of my own but I lost 10 hours of work

[werdwerdus]

well I'm using win xp x64 for my i3 system, but about a year or so ago i used nlite to strip down the os and make a new cd. it runs fine but there are a few problems with it such as i can't see the system with BoincView, firewall not working properly etc. but i have seemed to lost the original iso and cd for win xp x64. does someone have it that could send it to me? i have my own license keys so don't need to worry on that line

[trn]

After being off the grid for a week or two i'm back to crunch some WU's. I installed Win XP64 on a WCG/GPUGrid cruncher with a gtx460 in it. I had no plan on using swan_sync or leaving a thread open; I was expecting to take a small performance hit as I did with Win7. I was shocked, Win XP64 apparntly can't GPU crunch a GPUGrid WU without a freecore or swan_sync. Times to complete were showing at ~38 hours after 28 hours and only being 75% done on a KASHIF WU! I was expecting better from Win XP, I ran full 100% WCG on all cores with Win7 and only took a small hit with GPUGrid performance. I freed up a core and enabled swan_sync to double check if that was the problem. I might have to swap my GTX460 back into a win7 machine because i'm not sure I can take loosing a core for a GPU project Projected time to finish is just falling off my WU I type this now... I also got around to RMA'ing my broken gtx460 so i'm hoping to have a 2nd 460 for the grid also

[Snow Crash]

You sure everything is good with that card? ... sometimes a crash will degrade it to running at 2D performance level and a reboot is required to get things back into shape again. I ask because GPUGrid WUs always process faster on XP (32 or 64) than Win7 with or without SWAN_SYNC.

[trn]

You sure everything is good with that card? ... sometimes a crash will degrade it to running at 2D performance level and a reboot is required to get things back into shape again. I ask because GPUGrid WUs always process faster on XP (32 or 64) than Win7 with or without SWAN_SYNC.

No i'm not really sure, but everything looked OK to me except the GPUGrid WU times My copy of Win XP 64 is a little bit less than legit than I would like. Until I just turned on swan_sync the time to complete was showing at 12+ hours; that was 20 minutes ago. Now its down to under 6 hours.

*edit, even before swan_sync gpu usage showed up at 98 or 99% in afterburner. Much, much better than win7 ever did.

[CaptMorgan]

Yo trn- what you using there Ultra VNC?

[trn]

Yeah, Ultra VNC

[Snow Crash]

trn ... that WU restarted 4 times ... I know you are a serious cruncher so I hesitate to ask but ... you sure it didn't get turned off somehow?

Ultra VNC FTW

[trn]

I'm thinking that this less than legit Win XP x64 copy has issues. (supreme ub3r h4x0r3d w1n XP x64!!! or something lame like that.) And as far as GPUGrid goes i'm pretty much a noob I'll try a driver re-install before trashing this rubbish copy of XP, Its errored out on 17 GPUGrid WU's now! The computer is running stable, WCG has no errors.

And I'm like UltraVNC service and client, its better than windows remote desktop (despite the obvious that windows remote shuts off the GPU) because you can actually restart or power off the connected computer using the start menu which is very very helpful. Theres a little GUI lag, but its worth the trade-off for better features.

[roki977]

I had to RMA my 3x2 ddr3 kit and installed 3x1 kit and win7 32bit, before it was win7 64. Good thing is that GPU usage is higer now, 70-80%. With 64bit it was 60-72%. I can score over 50k per day with my gtx470 now, before it was 45k.

[Snow Crash]

Hey Roki ... glad to hear your rig is fixed
I doubt the DRAM is making a difference on GPUGrid,
it is more likely the different types of WUs but if you are now stable you will see better points

[roki977]

Not ram, but maybe 32bit win7. 58k yesterdy

[werdwerdus]

I think we are due to an update seeing as it's been over 2 years since the last post

Currently to get the best return rate (and consequently highest PPD) you really only need to concern yourself with a few simple things. Most of this is coming straight from my recent personal experience helped along by Snow Crash and some other users at the GPUGRID forums (mainly the first 3 items are my experience, to a lesser extent for some of the later points)


1. Set the appropriate task length for the cards in your machine. According to current FAQ http://www.gpugrid.net/forum_thread....rap=true#21666, there are 2 choices: short or long. In order to qualify for a 50% points bonus, the task needs to be reported within 24 hours, and within 48 hours for a 25% bonus. Current long tasks can take well over 48 hours to complete on GT 240 or lesser cards. For comparison my GTX 650 is averaging ~18-20 hours for long tasks.

2. Make sure your cache is set low enough. You don't want to download a new task when you still have several hours remaining on the task in progress. Consider a cache of between 0.05 and 0.3 days in BOINC preferences. This will help return completed tasks within the 24 or 48 bonus period which starts as soon as the task is downloaded to your machine. This is practically a non-issue for current gen fast cards though, since the project only allows 2 tasks per GPU regardless of cache setting. I have found that a low cache I.E. 0.01 may sometimes result in idle time between uploading a completed task and downloading the next task.

3. Set up BOINC to report completed tasks immediately. Create or modify cc_config.xml to include the setting. Read the post http://www.xtremesystems.org/forums/...=1#post4493060 if you are unsure. Disregard the section about SWAN_SYNC in that post. If you do not set up to report immediately, the reporting could be delayed several hours waiting for the project server to connect to your machine for results.


From what I can tell, these 3 items should give you nearly the maximum output for your system. To squeeze out the last few percent there are some other things to do.

There is no longer any need or use to enable the SWAN_SYNC variable. This is handled internally by the application (acemd).

Keeping 1 free core (physical or virtual/HT) per gpu is still recommended AFAIK, and can increase GPU utilization a few percent depending on the system hardware.

Depending on your specific GPU it may return results faster using an older driver (specificaly from the CUDA version used). From what I can tell this would apply to any GPUs less than Fermi but I would like some clarification.

I hope this is helpful to at least 1 person currently, but at least it may help for any new users joining our team.

[Snow Crash]

Thanks for updating this!

Depending on your specific GPU it may return results faster using an older driver (specificaly from the CUDA version used). From what I can tell this would apply to any GPUs less than Fermi but I would like some clarification.

Exactly correct. For 2xx series cards use a 285.x version driver. The 295 and 296 drivers have a low power bug and should be avoided.

[Santas]

how much better is windows XP then win7?? i could have xp on a old spin drive if its really worth it.

[fuzelogic]

I have 2 Nvidia cards in my system. 1 - GTX 285 and 1 - GTX 680 (recently purchased)

Before with only the GTS-295, my system draw when measure was around 450-watts

Now, when I have both cards running crunching including the processor, the UPS reads I'm pulling around 680 watts max.

I seem to have many Nvidia driver crashes and I'm not sure if it's due to the work unit or my configuration. The cards are not SLI or overclocked.

Does anyone else get alot of bluescreens? Sometimes when the driver crashes the screen will flicker a couple times but return back to the desktop after the driver restarts.

[werdwerdus]

I suggest to do a completely clean driver install. You will first need to do a driver sweep / driver cleaning and then reinstall the latest driver. What os do you use?

Any way it may be the GTX 285 using the cuda 4.2 app causing instability, which you cannot avoid if you are using a 600 series card with a 200 series card. You could try removing the gtx 285, doing a clean reinstall of the drivers, and see if there are still any problems with only the 680 installed.

Let us know what you try and the results.

[Kingcarcas]

Crap i hadn't even bothered looking into short or long, thanks!