Welcome to MilkyWay@home

GPU crunching question please


Advanced search

Message boards : Number crunching : GPU crunching question please
Message board moderation

To post messages, you must log in.

Previous · 1 · 2 · 3 · 4 · 5 · Next

AuthorMessage
ProfileUBT-Timby

Send message
Joined: 8 Oct 07
Posts: 3
Credit: 38,987,774
RAC: 0
30 million credit badge14 year member badge
Message 14292 - Posted: 7 Mar 2009, 23:36:46 UTC

Sorry guys if this is in the wrong place or the question has been asked before but .....

Nice new 4870 fitted in an AMD dual core PC installed the BM 6.4.5 and 8.12 drivers and off it goes processing 4 wu at a time, brilliant I thought, lets take it out and put it into a nice AMD quad machine with a Phenom 9850 cpu. Installed it exactly the same way as before with the same BM and driver set and it will only process 3 wu at once? Strange thing is that it say there are 8 wu running at once but only 3 progress. Can any one throw some light on this for me please?

Mark
ID: 14292 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Cluster Physik

Send message
Joined: 26 Jul 08
Posts: 627
Credit: 94,940,203
RAC: 0
50 million credit badge14 year member badgeextraordinary contributions badge
Message 14293 - Posted: 7 Mar 2009, 23:47:20 UTC - in response to Message 14292.  
Last modified: 8 Mar 2009, 0:03:48 UTC

Sorry guys if this is in the wrong place or the question has been asked before but .....

Nice new 4870 fitted in an AMD dual core PC installed the BM 6.4.5 and 8.12 drivers and off it goes processing 4 wu at a time, brilliant I thought, lets take it out and put it into a nice AMD quad machine with a Phenom 9850 cpu. Installed it exactly the same way as before with the same BM and driver set and it will only process 3 wu at once? Strange thing is that it say there are 8 wu running at once but only 3 progress. Can any one throw some light on this for me please?

Mark

You probably used version 0.19/b before and have now 0.19d. That behaviour is pefectly normal. You should have a look to the supplied readme file. It's all explained in quite some length there ;)
ID: 14293 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Cluster Physik

Send message
Joined: 26 Jul 08
Posts: 627
Credit: 94,940,203
RAC: 0
50 million credit badge14 year member badgeextraordinary contributions badge
Message 14294 - Posted: 8 Mar 2009, 0:02:48 UTC - in response to Message 14285.  
Last modified: 8 Mar 2009, 0:04:15 UTC

Ice, I've got the 3870 installed, Catalyst 9.1 drivers, no other application running, only Boinc with the optimized apps for ATI GPU (0.19d Win64), but it's crashing the graphics driver very often (the Catalyst Control Center is not running in the background). I can't leave my computer crunching away without checking up on it from time to time. Is it the nature of the beast as it's not a final version? When it does work it's very fast, making my Dell AMD dual core nearly beat my Mac Pro with 8 cores!

Additional to the things Ice mentioned you can try to play with the options in the app_info.xml (look to the readme file). Limiting the app to one concurrent WU (option n1) and/or increasing the wait factor can also help if the system can't stand the permanent high GPU load.
ID: 14294 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ProfileUBT-Timby

Send message
Joined: 8 Oct 07
Posts: 3
Credit: 38,987,774
RAC: 0
30 million credit badge14 year member badge
Message 14295 - Posted: 8 Mar 2009, 0:09:32 UTC - in response to Message 14294.  

Thanks for the swift reply Cluster Physik, I did notice the readme and have had a quick tinker with the app_info file. I will have a better look tomorrow as it late now, many thanks again

Mark
ID: 14295 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ProfilePaul D. Buck

Send message
Joined: 12 Apr 08
Posts: 621
Credit: 161,934,067
RAC: 0
100 million credit badge14 year member badge
Message 14308 - Posted: 8 Mar 2009, 2:21:33 UTC

Isn't there a law against reading a READ.ME file?

I could have sworn that there was a tremor when I even clicked on the file name ...
ID: 14308 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ProfileAlberto Tanikawa

Send message
Joined: 29 Jan 09
Posts: 32
Credit: 1,962,668
RAC: 0
1 million credit badge13 year member badge
Message 14317 - Posted: 8 Mar 2009, 7:05:37 UTC - in response to Message 14288.  

Install the 8.12 drivers ...


I'll try that once I have access to the Dell again.
ID: 14317 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ProfileGalaxyIce
Avatar

Send message
Joined: 6 Apr 08
Posts: 2018
Credit: 100,142,856
RAC: 0
100 million credit badge14 year member badge
Message 14320 - Posted: 8 Mar 2009, 7:33:00 UTC - in response to Message 14317.  
Last modified: 8 Mar 2009, 7:34:27 UTC

Install the 8.12 drivers ...


I'll try that once I have access to the Dell again.

I found that it was best to go into Control Panel, Add or Remove Programs, and remove the ATI display drivers first. The 0.19* apps seemed to behave better than with just adding 8.12 without removing existing ATI drivers first.

ID: 14320 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ProfileAlberto Tanikawa

Send message
Joined: 29 Jan 09
Posts: 32
Credit: 1,962,668
RAC: 0
1 million credit badge13 year member badge
Message 14325 - Posted: 8 Mar 2009, 8:30:08 UTC - in response to Message 14320.  

I found that it was best to go into Control Panel, Add or Remove Programs, and remove the ATI display drivers first. The 0.19* apps seemed to behave better than with just adding 8.12 without removing existing ATI drivers first.


Thanks again Ice!

Additional to the things Ice mentioned you can try to play with the options in the app_info.xml (look to the readme file). Limiting the app to one concurrent WU (option n1) and/or increasing the wait factor can also help if the system can't stand the permanent high GPU load.


I did play with the n option a bit, tried 2 and 4, but left it at 2. I didn't change the wait factor though.
ID: 14325 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ProfileZanth
Avatar

Send message
Joined: 18 Feb 09
Posts: 158
Credit: 110,630,897
RAC: 1,642
100 million credit badge13 year member badge
Message 14383 - Posted: 8 Mar 2009, 18:17:00 UTC - in response to Message 14325.  

I have a little story, so bear with me before I get to my actual question.

CPU: Intel Core2Quad Q6600 OC'd to 3 GHz
GPU: Radeon HD3870

I just read up on the 19d app, souded good so I thought I'd give it a go. Running Windows 7 on Vista 64 Catalyst 9.2 with proper ati*.dll files in my sys32 folder. It ran pretty well running its 3 WUs in way better time than the 19 app did. Then... the driver crashed. Restarted, did it again, of course. So I read up more on the xml file and tweaked a bit, first to n1, and the GPU still ran at max load constantly, so I started messing with the w variable, and had to get it up to 1.6 before the load decreased. But by now, my crunching time had suffered so much it was worse off than the 19 app. So I found this here thread and saw that the 8.12 drivers are supposed to work way better. After fighting with Windows 7 for about 2 hours, I FINALLY get the install to go through. Cautiously, I started at n2 and no w variable. About 1 minute in display crash again. So I hit up n1 and it's going well. Then I had this thought, I have an extra 4850 laying around(the 3870 has better cooling as well as GDDR4 for gaming performance(, so why not toss that in just to crunch since the 19d app supports multiple GPUs. Load that up, strongarm windows into using the right drivers and fire boinc back up. Still, only one WU crunches. It was my understanding, and please, if I'm flat out wrong, I apologize, that the n1 flag would let one WU crunch per gpu? If this is the case, what might I be doing wrong?
ID: 14383 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ProfileGalaxyIce
Avatar

Send message
Joined: 6 Apr 08
Posts: 2018
Credit: 100,142,856
RAC: 0
100 million credit badge14 year member badge
Message 14390 - Posted: 8 Mar 2009, 18:31:14 UTC - in response to Message 14383.  
Last modified: 8 Mar 2009, 18:33:55 UTC

I have a little story, so bear with me before I get to my actual question.

CPU: Intel Core2Quad Q6600 OC'd to 3 GHz
GPU: Radeon HD3870

I just read up on the 19d app, souded good so I thought I'd give it a go. Running Windows 7 on Vista 64 Catalyst 9.2 with proper ati*.dll files in my sys32 folder. It ran pretty well running its 3 WUs in way better time than the 19 app did. Then... the driver crashed. Restarted, did it again, of course. So I read up more on the xml file and tweaked a bit, first to n1, and the GPU still ran at max load constantly, so I started messing with the w variable, and had to get it up to 1.6 before the load decreased. But by now, my crunching time had suffered so much it was worse off than the 19 app. So I found this here thread and saw that the 8.12 drivers are supposed to work way better. After fighting with Windows 7 for about 2 hours, I FINALLY get the install to go through. Cautiously, I started at n2 and no w variable. About 1 minute in display crash again. So I hit up n1 and it's going well. Then I had this thought, I have an extra 4850 laying around(the 3870 has better cooling as well as GDDR4 for gaming performance(, so why not toss that in just to crunch since the 19d app supports multiple GPUs. Load that up, strongarm windows into using the right drivers and fire boinc back up. Still, only one WU crunches. It was my understanding, and please, if I'm flat out wrong, I apologize, that the n1 flag would let one WU crunch per gpu? If this is the case, what might I be doing wrong?

Sounds a familiar to the problem I was getting (when 9.2 worked for a while then kept crashing). My theory is that you need to uninstall 9.2 completely - including the CCC Catalyst Control Centre. Otherwise you may have settings somewhere in your system that is causing an overclock with 8.12.

Flush it all out with a complete uninstall of all ATI products before installing 8.12 - is what I would advise.

ID: 14390 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ProfileZanth
Avatar

Send message
Joined: 18 Feb 09
Posts: 158
Credit: 110,630,897
RAC: 1,642
100 million credit badge13 year member badge
Message 14395 - Posted: 8 Mar 2009, 18:40:57 UTC - in response to Message 14390.  

Oh, my apology, I did omit a step. I uninstalled 9.2, wiped it clean with Driver Cleaner Pro.
ID: 14395 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ProfileGalaxyIce
Avatar

Send message
Joined: 6 Apr 08
Posts: 2018
Credit: 100,142,856
RAC: 0
100 million credit badge14 year member badge
Message 14397 - Posted: 8 Mar 2009, 18:45:48 UTC - in response to Message 14395.  

Oh, my apology, I did omit a step. I uninstalled 9.2, wiped it clean with Driver Cleaner Pro.

Then in that case I would try 9.1


ID: 14397 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ProfileZanth
Avatar

Send message
Joined: 18 Feb 09
Posts: 158
Credit: 110,630,897
RAC: 1,642
100 million credit badge13 year member badge
Message 14398 - Posted: 8 Mar 2009, 18:49:33 UTC - in response to Message 14397.  

Well, it's been running well since I set it to n1, crashed once when power saver turned the display off. The biggest thing I am curious about at the moment, is: I have a 3870 and a 4850 in it at the moment, but with n1, only one WU is running... shouldn't there be two?

And btw, thank you very much fior taking the time to help me out.
ID: 14398 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ProfileGalaxyIce
Avatar

Send message
Joined: 6 Apr 08
Posts: 2018
Credit: 100,142,856
RAC: 0
100 million credit badge14 year member badge
Message 14408 - Posted: 8 Mar 2009, 19:19:01 UTC - in response to Message 14398.  

Well, it's been running well since I set it to n1, crashed once when power saver turned the display off. The biggest thing I am curious about at the moment, is: I have a 3870 and a 4850 in it at the moment, but with n1, only one WU is running... shouldn't there be two?

The optimim is n3. n1 will only run one work unit at a time. I have mine set to n2, but with my 4850 4 WUs are always crunching at once. With my 4870, just 2 crunch as expected.

And btw, thank you very much fior taking the time to help me out

no problem :)


ID: 14408 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ProfileZanth
Avatar

Send message
Joined: 18 Feb 09
Posts: 158
Credit: 110,630,897
RAC: 1,642
100 million credit badge13 year member badge
Message 14410 - Posted: 8 Mar 2009, 19:27:57 UTC - in response to Message 14408.  

Ok, so its not something I'm doing wrong, the flag just isn't operating as expected on my machine. n1 should run 2 WUs, one per GPU, by the wording in the readme anyway. :) It's all good, I just wanted to sort out if I messed something, or if it was just a bit glitchy yet. I removed the n flag all together now and disabled power saving and its been running good so far, both GPUs are showing 99.9% utilization. :)
ID: 14410 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ProfileGalaxyIce
Avatar

Send message
Joined: 6 Apr 08
Posts: 2018
Credit: 100,142,856
RAC: 0
100 million credit badge14 year member badge
Message 14415 - Posted: 8 Mar 2009, 19:47:55 UTC - in response to Message 14410.  
Last modified: 8 Mar 2009, 19:48:37 UTC

...its been running good so far, both GPUs are showing 99.9% utilization. :)

That's good. But what I'm trying to do is to run another project with the spare CPU that the GPU isn't using. I'm succeeding so far with the 4870 and have a CPDN task running. If I get two CPDN, or one CPDN and a Cosmology, the MW crunching stops.

As for the 4850 - it's still cracking on with 4 MW tasks, but doesn't seem to want to share the CPU.

ID: 14415 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ProfileZanth
Avatar

Send message
Joined: 18 Feb 09
Posts: 158
Credit: 110,630,897
RAC: 1,642
100 million credit badge13 year member badge
Message 14417 - Posted: 8 Mar 2009, 20:06:45 UTC - in response to Message 14415.  

Mine has now crashed twice in rapid succession. :/ I think I'm going to chalk it up to Windows 7.
ID: 14417 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Brickhead
Avatar

Send message
Joined: 20 Mar 08
Posts: 108
Credit: 2,562,515,681
RAC: 0
2 billion credit badge14 year member badgeextraordinary contributions badge
Message 14422 - Posted: 8 Mar 2009, 20:45:50 UTC

To have the desired number of MW WUs run alongside the desired number of other projects' WUs, you need to play with the resource share. IME, that's the only way, as the avg_ncpus parameter has no influence on that whatsoever. Furthermore, I see no reason to allocate a whole core to the GPU app, it needs much less. I'll try to explain my findings - and my humble recommendations:

How many WUs to run simultaneously on each GPU?

command: n2

How many CPU cores to let the GPU app use at max?

max_ncpus: 1

How many GPU WUs to allow while still running CPU WUs (from other projects) on all cores?

avg_ncpus: less than 1 / (n * ngpu)
(nevermind, just set it to 0.1)

How many GPU WUs to actually run at once?

BOINC resource share: n * ngpu / (n * ngpu + ncpu)
n2, one gpu, dual-core cpu -> 50%
n2, one gpu, quad-core cpu -> 33%
n2, two gpus, dual-core cpu -> 67%
n2, twp gpus, quad-core cpu -> 50%

(Of course, the 'n' reference only applies to 0.19d so far.)

On my quad-core with a (dual-GPU) 4870x2, I have 4 MW WUs running alongside 4 WUs from other projects - shown as 8 WUs all in 'running' state.
On my quad-core with a (single-GPU) 4870/1g, I have BOINC thinking that 4 MW WUs run alongside 4 WUs alongside 4 WUs from other projects, but the GPU app has 2 of them on hold thanks to the n2 command.

It takes a while for the resource share to 'settle down', and the CC scheduler will act up slightly now and then, but in genereal, it's as simple as getting the maths straight and give it a few hours.
ID: 14422 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ProfileGalaxyIce
Avatar

Send message
Joined: 6 Apr 08
Posts: 2018
Credit: 100,142,856
RAC: 0
100 million credit badge14 year member badge
Message 14429 - Posted: 8 Mar 2009, 21:03:01 UTC - in response to Message 14422.  

<snip>

... it's as simple as getting the maths straight and give it a few hours.

Fantastic! Well done Brickhead!, that's done the trick for me :D


ID: 14429 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Brickhead
Avatar

Send message
Joined: 20 Mar 08
Posts: 108
Credit: 2,562,515,681
RAC: 0
2 billion credit badge14 year member badgeextraordinary contributions badge
Message 14441 - Posted: 8 Mar 2009, 21:17:06 UTC - in response to Message 14429.  

Fantastic! Well done Brickhead!, that's done the trick for me :D

Uh-oh, does this mean I'd better watch my back now? :D

Happy to learn that my findings still hold water now that we have a statistical base of no less than 4 :P
ID: 14441 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Previous · 1 · 2 · 3 · 4 · 5 · Next

Message boards : Number crunching : GPU crunching question please

©2022 Astroinformatics Group