GPU RAM: any ideas on how to free up a few MB?


log in

Advanced search

Questions and Answers : GPU applications : GPU RAM: any ideas on how to free up a few MB?

Author Message
Brendan Burwood
Send message
Joined: 1 Jul 99
Posts: 4
Credit: 707,694
RAC: 0
Australia
Message 980371 - Posted: 18 Mar 2010, 4:25:08 UTC

I've recently installed BOINC on a cuda capable machine (after a complete reinstall) with an 8400GS 256MB card (running the nVidia 196.75 drivers). BOINC dutifully downloaded over 1000 cuda work units and then left them "Waiting to start", while running 2 CPDN units normally. Turns out the problem is in the amount of *free* GPU RAM: I was getting the dreaded "231MB < 238MB" message and the cuda tasks refused to run. The only way so far I can get them to run is by dropping the resolution (at present I've got it at 1024x768x32bit) while the pc is not being used, so that it at least does a few units today, but this is NOT a long term solution - it can only ever be a temporary workaround. The machine has a 1680x1050 LCD screen so that's the resolution I need to have it running in for obvious reasons.

So ... 2 things then:
1. Does anyone have any ideas on how to free up 8MB of GPU RAM *WITHOUT* dropping the resolution from the native LCD values above? (dropping to 16bit or even 8bit is still not quite enough for SETI and only frees up about 5 and 7MB respectively)

2. Is there any way for SETI CUDA to correctly report to BOINC exactly how much *free* GPU RAM it requires rather than taking the reported size of GPU RAM as OK. Or, is there an optimised SETI CUDA version that requiers a little less GPU RAM than the standard version? (perhaps requiring <230MB or <225MB to be safe - maybe the SETI developers might see if they could perhaps release a new version requiring a little less GPU RAM for those of us with borderline amounts free)

bb from Oz.
____________

Profile Ageless
Avatar
Send message
Joined: 9 Jun 99
Posts: 12396
Credit: 2,666,323
RAC: 958
Netherlands
Message 980401 - Posted: 18 Mar 2010, 6:07:08 UTC - in response to Message 980371.
Last modified: 18 Mar 2010, 6:37:18 UTC

running the nVidia 196.75 drivers

These do that. Those are CUDA 2.3/CUDA 3.0 drivers, which activate the cuda23 app_plan, which require that you have a minimum of 234 (or 231) MB free on the card. Before that it didn't matter much what resolution you ran in, as even on a 1920x1200x32bpp resolution it would consume only 10MB of VRAM, leaving you with ~245MB videomemory free.

Since an 8400 GS is never a true CUDA 2.3 or above GPU, you can easily go back to an earlier driver.

Later edit: Although, if you want to use BOINC 6.10, you will need a minimum of 185.85
____________
Jord

Fighting for the correct use of the apostrophe, together with Weird Al Yankovic

Brendan Burwood
Send message
Joined: 1 Jul 99
Posts: 4
Credit: 707,694
RAC: 0
Australia
Message 980411 - Posted: 18 Mar 2010, 6:45:41 UTC - in response to Message 980401.

running the nVidia 196.75 drivers

These do that. Those are CUDA 2.3/CUDA 3.0 drivers, which activate the cuda23 app_plan, which require that you have a minimum of 234 (or 231) MB free on the card. Before that it didn't matter much what resolution you ran in, as even on a 1920x1200x32bpp resolution it would consume only 10MB of VRAM, leaving you with ~245MB videomemory free.

Since an 8400 GS is never a true CUDA 2.3 or above GPU, you can easily go back to an earlier driver.

Later edit: Although, if you want to use BOINC 6.10, you will need a minimum of 185.85
Now that is interesting. In that case I'll drop back to a 19x.xx driver later tonight. I have 190.62 lying around, sounds like that should work ok. ... and they tell you that updated drivers are better! Not in this case apparently.

I'll post again after the "downgrade" to say how things went.

bb from Oz.
____________

Profile Ageless
Avatar
Send message
Joined: 9 Jun 99
Posts: 12396
Credit: 2,666,323
RAC: 958
Netherlands
Message 980416 - Posted: 18 Mar 2010, 7:26:44 UTC - in response to Message 980411.

One thing I omitted (was in time constraints, Dirty Jobs was about to start on Discovery ;-)):

177.35 is the ultimate minimal driver you need to run CUDA with on Seti. That's CUDA 2.0 (For BOINC 6.4.5 to 6.6.41, the clients from 6.6 onwards will check for this driver version to be the minimum)

185.85 is CUDA 2.2 (For BOINC 6.10)

190.38 is CUDA 2.3 (and activates the cuda23 app_plan in the client and on the server)

195.39 and above is CUDA 3.0

I hope these help you muddle through it all. :-)
____________
Jord

Fighting for the correct use of the apostrophe, together with Weird Al Yankovic

Profile BilBg
Volunteer tester
Avatar
Send message
Joined: 27 May 07
Posts: 2879
Credit: 6,455,094
RAC: 2,765
Bulgaria
Message 980666 - Posted: 18 Mar 2010, 22:54:36 UTC - in response to Message 980411.
Last modified: 18 Mar 2010, 23:44:25 UTC

If you install
Lunatics' Unified Installer v0.2
http://setiathome.berkeley.edu/forum_thread.php?id=54288
http://lunatics.kwsn.net/index.php?module=Downloads;catd=9

it will install CUDA 2.2 DLLs and use 6.08 opt app (with any NVIDIA driver > 185.85; 190.xx & 191.xx are faster than 195.xx)
Requires 220 MB free Video RAM

You will be able to use ReSchedule 1.9 to move the WUs from CPU to GPU and vice versa:
http://setiathome.berkeley.edu/forum_thread.php?id=59291


Important:
Upload and Report any completed CUDA work and exit BOINC before install -
all CUDA WUs branded 6.09 will be aborted (deleted) and new 6.08 WUs will be downloaded after you restart BOINC.

.
____________



- ALF - "Find out what you don't do well ..... then don't do it!" :)

Brendan Burwood
Send message
Joined: 1 Jul 99
Posts: 4
Credit: 707,694
RAC: 0
Australia
Message 980718 - Posted: 19 Mar 2010, 1:33:26 UTC

I would have posted sooner, but my cable was flakey - fixed now so it should be fine for another 10 years. Water + electronics don't play nicely togther, but lasted surprisingly well! The techie opened the box on the overhead cable on the street and poured some water out of it, hehe! (and it's a fine sunny day today) New box and parts now. :-)

Thanks for the extra info, but I tried 186.18 and free GPU RAM dropped back to 221MB! So I've put the new 197.13 WHQL candidate driver on and it's back up to 232MB now. I guess that means I need to look at the optimised aps now if I want to use CUDA. With 220MB required for those I should hopefully be fine.

... and yes, I'm using BOINC 6.10.36.

bb from Oz.
____________

Brendan Burwood
Send message
Joined: 1 Jul 99
Posts: 4
Credit: 707,694
RAC: 0
Australia
Message 980744 - Posted: 19 Mar 2010, 2:47:52 UTC
Last modified: 19 Mar 2010, 2:48:43 UTC

All done and appearing to be running nicely with the optimised apps, although the time estimated with the optimised versions appears to be very wrong, as in grossly overestimated. Any way to get it to give a more accurate estimate?

The stock cuda version (when I got it to run by dropping the screen res) was taking around 105-110 minutes for the shorter units and estimating around 5.5 hours for the longer units, the new version is estimating 5 hours for the shorter units and 18 hours for the longer units, but does appear to be on track to do a short unit in around 95-100 minutes, so it is a little quicker 0 but at least it is running with native LCD res at full 32bit colour. :-)

So now it's crunching its way though 1000+ units 3 at a time (2 on cpu 1 on gpu) in panic mode. It probably won't have enough time to do them all before the deadline, but we'll see how far it gets. I have also set the ReSchedule tool to run when windows starts, and check every 6 hours. (with the 2 checkboxes on the left of the settings cleared) CPDN has been put on the backburner as a result, but I'm sure it will get its share of cpu time soon.

I guess I should now tell boinc to use the gpu based on preference, which I will set to either 1 minute or 0.5 minutes before using the gpu ... otherwise I might get complaints for the machine being slow, hehe.

Thanks for all the pointers in the right direction.

bb from Oz.
____________

Profile perryjay
Volunteer tester
Avatar
Send message
Joined: 20 Aug 02
Posts: 3377
Credit: 16,210,850
RAC: 5,873
United States
Message 980751 - Posted: 19 Mar 2010, 3:08:38 UTC - in response to Message 980744.

Brenden,
Glad you got things going. As to getting a little more accurate estimate try the links in this post http://setiathome.berkeley.edu/forum_thread.php?id=59270&nowrap=true#980502 Between them they explain how to figure your GFlops and where to put them in your app_info.
____________


PROUD MEMBER OF Team Starfire World BOINC

Profile BilBg
Volunteer tester
Avatar
Send message
Joined: 27 May 07
Posts: 2879
Credit: 6,455,094
RAC: 2,765
Bulgaria
Message 981557 - Posted: 20 Mar 2010, 8:16:39 UTC - in response to Message 980744.
Last modified: 20 Mar 2010, 8:25:39 UTC

Time estimate will become more and more accurate after each completed WU
as BOINC have to learn how fast are the new opt apps on your hardware.

On CPU they are usually 150% - 250% faster, on GPU - 105% - 120% (for the same CUDA version)

You can try CUDA 2.3 DLLs to see will they work on your GPU / Video RAM
(get them from Lunatics, keep the "old" 2.2 DLLs in case you have to revert)
As always Report before making changes.


P.S.
Sometimes ReSchedule do not work correct in auto mode

.
____________



- ALF - "Find out what you don't do well ..... then don't do it!" :)

Questions and Answers : GPU applications : GPU RAM: any ideas on how to free up a few MB?

Copyright © 2014 University of California