Advanced search

Message boards : News : CUDA upgrade notice

Author Message
Profile GDF
Volunteer moderator
Project administrator
Project developer
Project tester
Volunteer developer
Volunteer tester
Project scientist
Send message
Joined: 14 Mar 07
Posts: 1895
Credit: 629,356
RAC: 0
Level
Gly
Scientific publications
watwatwatwatwat
Message 27203 - Posted: 3 Nov 2012 | 11:24:40 UTC
Last modified: 4 Nov 2012 | 11:42:33 UTC

By beginning of next year, we would like to remove the cuda3.1 applications.
These are still the old apps and don't have several of the new features available in the new app which we need to do the science.

We will probably have a cuda4.2 app only, so within the next few months, drivers should be upgraded to at least this version. In our machines, the latest driver 304.51 works quite nicely under Linux, very robust.

gdf

Wiyosaya
Send message
Joined: 22 Nov 09
Posts: 111
Credit: 171,819,453
RAC: 392,567
Level
Ile
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 27204 - Posted: 3 Nov 2012 | 12:22:00 UTC - in response to Message 27203.

Please do. Due to their inefficiency, the sooner the better, IMHO.
____________

tomba
Send message
Joined: 21 Feb 09
Posts: 497
Credit: 700,690,702
RAC: 0
Level
Lys
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 27205 - Posted: 3 Nov 2012 | 16:59:45 UTC
Last modified: 3 Nov 2012 | 17:17:24 UTC

Pretty please, for the benefit of those of us who know not what CUDA 3.1 and CUDA 4.2 are, all we need to know is what BOINC and Nvidia drivers we should be using (I think...).

Also, it would be good to know well in advance when you plan to drop support for intermediate cards, e.g., my faithful GTX 460., which I bought just for GPUGRID when you put my GSO 9600 out to pasture.

ALSO - If Nathan is still with you, he who I assume is responsible for NATHAN WUs, can he please stop sending me WUs that totally clog my PC? If he wants to know which WUs, I have that information - I've been keeping a log...

The log tell me that, of the 30 NATHANs I've had these past three weeks, 50% I've aborted so I can get a few minutes of peaceful, quiet PC time, just for me, not for NATHAN!!

Thank you.
____________

tomba
Send message
Joined: 21 Feb 09
Posts: 497
Credit: 700,690,702
RAC: 0
Level
Lys
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 27207 - Posted: 3 Nov 2012 | 19:37:37 UTC

In the last hour I've aborted four NATHANs for killing my PC performance. This is getting ridiculous!!
____________

Profile GDF
Volunteer moderator
Project administrator
Project developer
Project tester
Volunteer developer
Volunteer tester
Project scientist
Send message
Joined: 14 Mar 07
Posts: 1895
Credit: 629,356
RAC: 0
Level
Gly
Scientific publications
watwatwatwatwat
Message 27211 - Posted: 4 Nov 2012 | 11:43:38 UTC - in response to Message 27207.

nathan WU should have no difference from the others if not that they are large systems.
What is the problem exactly?


gdf

Profile skgiven
Volunteer moderator
Project tester
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,834,906,524
RAC: 276,970
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 27213 - Posted: 4 Nov 2012 | 15:45:03 UTC - in response to Message 27205.

Pretty please, for the benefit of those of us who know not what CUDA 3.1 and CUDA 4.2 are, all we need to know is what BOINC and Nvidia drivers we should be using (I think...).

Also, it would be good to know well in advance when you plan to drop support for intermediate cards, e.g., my faithful GTX 460., which I bought just for GPUGRID when you put my GSO 9600 out to pasture.


Fermi cards can run CUDA 4.2 tasks. This includes the GTX 460.
Older cards (GeForce 200 series and below) may not be capable of running CUDA 4.3 and would not perform well.
____________
FAQ's

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

Profile dskagcommunity
Avatar
Send message
Joined: 28 Apr 11
Posts: 456
Credit: 810,073,458
RAC: 0
Level
Glu
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwat
Message 27215 - Posted: 4 Nov 2012 | 16:16:33 UTC
Last modified: 4 Nov 2012 | 16:18:28 UTC

Yes when 31 get killed we can all throgh out our 2xx cards :/ but im happy that you let them actived until today. Im sure i have a other project in reserve for the 285 gtx when you do the final change ^^ i think yo will inform us a week before change? Perhaps i have enough money to buy a new nvidia card then after christmas ^^
____________
DSKAG Austria Research Team: http://www.research.dskag.at



Crunching for my deceased Dog who had "good" Braincancer..

Profile GDF
Volunteer moderator
Project administrator
Project developer
Project tester
Volunteer developer
Volunteer tester
Project scientist
Send message
Joined: 14 Mar 07
Posts: 1895
Credit: 629,356
RAC: 0
Level
Gly
Scientific publications
watwatwatwatwat
Message 27219 - Posted: 4 Nov 2012 | 18:29:18 UTC - in response to Message 27215.

We might be able to keep the old application for cuda3.1, but at least one of the two queues must be with just the new app.

gdf

werdwerdus
Send message
Joined: 15 Apr 10
Posts: 123
Credit: 1,004,473,861
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 27222 - Posted: 4 Nov 2012 | 23:50:23 UTC - in response to Message 27219.

We might be able to keep the old application for cuda3.1, but at least one of the two queues must be with just the new app.

gdf

please make the long run queue only 4.2+
____________
XtremeSystems.org - #1 Team in GPUGrid

Profile MarkJ
Volunteer moderator
Project tester
Volunteer tester
Send message
Joined: 24 Dec 08
Posts: 730
Credit: 189,243,545
RAC: 0
Level
Ile
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 27226 - Posted: 5 Nov 2012 | 11:00:07 UTC - in response to Message 27222.

We might be able to keep the old application for cuda3.1, but at least one of the two queues must be with just the new app.

gdf

please make the long run queue only 4.2+


I thought the original plan was to use cuda40 and cuda42. If we're going to keep cuda31 around then limit it to the short queue, with the beta and long queues using cuda42.

Is there any advantage to using cuda50 at this point or would it be too much development effort?
____________
BOINC blog

Profile GDF
Volunteer moderator
Project administrator
Project developer
Project tester
Volunteer developer
Volunteer tester
Project scientist
Send message
Joined: 14 Mar 07
Posts: 1895
Credit: 629,356
RAC: 0
Level
Gly
Scientific publications
watwatwatwatwat
Message 27231 - Posted: 5 Nov 2012 | 19:27:38 UTC - in response to Message 27226.

for cuda5 there is no point, no speed advantage. The problem of cuda4.0 is that it seems buggy.
We could do cuda4.1 instead of cuda4.2, but there is little change.

gdf

Profile skgiven
Volunteer moderator
Project tester
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,834,906,524
RAC: 276,970
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 27233 - Posted: 6 Nov 2012 | 9:55:34 UTC - in response to Message 27231.

For now could you place,
4.2 WU's into the Long queue only,
3.1 WU's into the short queue,
and if need be put shorter 4.2 WU's into the Beta queue?
____________
FAQ's

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

tomba
Send message
Joined: 21 Feb 09
Posts: 497
Credit: 700,690,702
RAC: 0
Level
Lys
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 27234 - Posted: 6 Nov 2012 | 19:35:26 UTC - in response to Message 27211.

nathan WU should have no difference from the others if not that they are large systems.
What is the problem exactly?gdf


Hello GDF. I shall be delighted to tell you EXACTLY what the problems are despite the fact that they have produced no solutions in
this thread.

I can tell you EXACTLY which NATANs kill my PC performance, and which don't.

If you're listening, I'm ready...

Tom
____________

flashawk
Send message
Joined: 18 Jun 12
Posts: 241
Credit: 1,692,519,522
RAC: 706,915
Level
His
Scientific publications
watwatwatwatwatwatwatwat
Message 27239 - Posted: 7 Nov 2012 | 6:52:52 UTC

You only have 6GB of memory with Windows 7 no less, it's amazing you can run any GPU wu's at all. You're getting alot of reads and writes to you're scratch file, that would make it diffacult for any computer to respond quickly. Anyway, just a thought with RAM so cheap these days.

Profile skgiven
Volunteer moderator
Project tester
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,834,906,524
RAC: 276,970
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 27241 - Posted: 7 Nov 2012 | 10:56:06 UTC - in response to Message 27239.
Last modified: 7 Nov 2012 | 13:06:55 UTC

6GB is normally enough - even for W7.
If tomba was running 8 Rosetta tasks then it might be an issue as those tasks can use 400BM RAM, so 3GB of RAM could be being used. More importantly, the CPU could be being saturated. W7 typically uses around 2GB RAM. Of course other RAM heavy apps (including browsers) could be an issue too, but tomba's problem is specific to one type of GPUGrid WU.
Anyway, none of this has anything to do with this thread, 'CUDA upgrade notice'.
____________
FAQ's

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

Profile dskagcommunity
Avatar
Send message
Joined: 28 Apr 11
Posts: 456
Credit: 810,073,458
RAC: 0
Level
Glu
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwat
Message 27247 - Posted: 7 Nov 2012 | 13:28:05 UTC - in response to Message 27239.

You only have 6GB of memory with Windows 7 no less, it's amazing you can run any GPU wu's at all. You're getting alot of reads and writes to you're scratch file, that would make it diffacult for any computer to respond quickly. Anyway, just a thought with RAM so cheap these days.


Lol then i must have massive problems on my 2gb ram machine with win7 ;) *ironic* 6gb is way enough O.o
____________
DSKAG Austria Research Team: http://www.research.dskag.at



Crunching for my deceased Dog who had "good" Braincancer..

tomba
Send message
Joined: 21 Feb 09
Posts: 497
Credit: 700,690,702
RAC: 0
Level
Lys
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 27249 - Posted: 7 Nov 2012 | 17:05:46 UTC - in response to Message 27241.

Anyway, none of this has anything to do with this thread, 'CUDA upgrade notice'.


You're absolutely right and I apologize, but I figured that this thread was started by someone on the development team who might, perhaps, bring my problem to the attention of their colleague who does NATHANs...

____________

sardine
Send message
Joined: 17 Sep 08
Posts: 4
Credit: 67,875
RAC: 0
Level

Scientific publications
wat
Message 27317 - Posted: 15 Nov 2012 | 23:00:41 UTC - in response to Message 27203.

I got a big problem. I can't complete any task on my computer.
I use asus GTS 450 with NVIDIA DRIVERS 301.42WHQL 、 intel E5200 、 4G DDRII (PC3200)、Windows 7 Enterprise (x64) and BOINC 7.0.28(x64).

but...
It's seems too bad >.<

[/code]

lukeu
Send message
Joined: 14 Oct 11
Posts: 19
Credit: 39,939,809
RAC: 9,806
Level
Val
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 27318 - Posted: 16 Nov 2012 | 12:34:07 UTC

I updated drivers for my GTX 260 yesterday, and cuda42 work units apparently started crashing.

Last 3 tasks (standard queue):
* 6044359 - ACEMD2: GPU molecular dynamics v6.16 (cuda42) - CRASHED
* 6047209 - ACEMD2: GPU molecular dynamics v6.16 (cuda31) - success
* 6048860 - ACEMD2: GPU molecular dynamics v6.16 (cuda42) - CRASHED

I had no problem prior to the update. (http://www.gpugrid.net/results.php?userid=81842). Indeed task 6044359 was running fine, then crashed immediately after restarting BOINC, following the driver update.

NB: The driver (306.97 WHQL) arrived via Windows Update yesterday. Prior to that, I was running the last version Windows Update sent me (295.73 I think, possibly 296.10). I wonder if others have starting seeing similar issues now that WU has pushed this driver?

GTX 260, Win7-x64, BOINC 7.0.28(x64), Intel i7-860

Profile dskagcommunity
Avatar
Send message
Joined: 28 Apr 11
Posts: 456
Credit: 810,073,458
RAC: 0
Level
Glu
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwat
Message 27319 - Posted: 16 Nov 2012 | 15:03:11 UTC
Last modified: 16 Nov 2012 | 15:05:20 UTC

yes because the 2xx series (i wrote too in this thread) is not good for cuda42..so get your driver back down in 2xx Range. Run it until end of the year and when it changes switch fully to short WUs where it seems perhaps to continue with 31 when i read the thread correct. Or buy a new card for gpugrid. The only chance is to downclock core (-50Mhz?) and memory (-100Mhz) from the graphiccard and look if it perhaps helps. but this tip is more for some fermi cards (like my 560TI) who produce cuda42 errors from time to time too.
____________
DSKAG Austria Research Team: http://www.research.dskag.at



Crunching for my deceased Dog who had "good" Braincancer..

werdwerdus
Send message
Joined: 15 Apr 10
Posts: 123
Credit: 1,004,473,861
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 27320 - Posted: 16 Nov 2012 | 16:44:11 UTC

I recommend to use driver 285 for 2xx cards
____________
XtremeSystems.org - #1 Team in GPUGrid

lukeu
Send message
Joined: 14 Oct 11
Posts: 19
Credit: 39,939,809
RAC: 9,806
Level
Val
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 27321 - Posted: 16 Nov 2012 | 17:10:38 UTC

Thanks guys, rolling back to 295.73 for my GTX 260 was surprising easy! (just a few clicks, no reboot) and GPUGrid is happily warming my toes once again.

The monitor-power bug in 295/296 shouldn't affect me.

Certainly thinking about a GTX 660! Although I hear it's not so good at heating... :-|

sardine
Send message
Joined: 17 Sep 08
Posts: 4
Credit: 67,875
RAC: 0
Level

Scientific publications
wat
Message 27335 - Posted: 18 Nov 2012 | 14:49:43 UTC - in response to Message 27317.

I have fixed the problem. HCC and GPUGRD can't work together.
I had removed the project HCC(Help Conquer Cancer) under WCG(World Community Grid), and finished a long rungs task in 30 hours..http://www.gpugrid.net/workunit.php?wuid=3841543

By the way, I also update my nvidia drivers version to 306.97WHQL(CUDA version 5.0). It works.

I got a big problem. I can't complete any task on my computer.
I use asus GTS 450 with NVIDIA DRIVERS 301.42WHQL 、 intel E5200 、 4G DDRII (PC3200)、Windows 7 Enterprise (x64) and BOINC 7.0.28(x64).

but...
It's seems too bad >.<

http://www.gpugrid.net/results.php?hostid=138448

Profile dskagcommunity
Avatar
Send message
Joined: 28 Apr 11
Posts: 456
Credit: 810,073,458
RAC: 0
Level
Glu
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwat
Message 27346 - Posted: 19 Nov 2012 | 11:49:07 UTC
Last modified: 19 Nov 2012 | 11:50:07 UTC

Is there a possibility to switch earlier? ^^ or are there more changes needed before that needs time?
____________
DSKAG Austria Research Team: http://www.research.dskag.at



Crunching for my deceased Dog who had "good" Braincancer..

Profile GDF
Volunteer moderator
Project administrator
Project developer
Project tester
Volunteer developer
Volunteer tester
Project scientist
Send message
Joined: 14 Mar 07
Posts: 1895
Credit: 629,356
RAC: 0
Level
Gly
Scientific publications
watwatwatwatwat
Message 27368 - Posted: 20 Nov 2012 | 11:29:37 UTC - in response to Message 27346.

we want to wait as long as possible such many people will be already using the new drivers.

gdf

Wiyosaya
Send message
Joined: 22 Nov 09
Posts: 111
Credit: 171,819,453
RAC: 392,567
Level
Ile
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 27499 - Posted: 2 Dec 2012 | 3:40:47 UTC

From my experience, 3.1 apps tend to take substantially more time to run than the same type of WU running with the 4.2 app.

Take, for instance, this 3.1 WU - a Nathan respawn http://www.gpugrid.net/workunit.php?wuid=3863083
and this 4.2 WU - also a Nathan respawn http://www.gpugrid.net/workunit.php?wuid=3846597

Note that both were on my 580 machine, granted the same amount of credit, and that the 3.1 WU took almost twice as long as the 4.2 WU. Personally, I find situations like this particularly frustrating since it seems extremely inefficient to run 3.1 apps on machines that support the 4.2 app.

I have no direct experience in what it would take to program a project server to behave in this manner, however, projects like Einstein, Milkyway, and now Help Conquer Cancer at World Community Grid are able to figure out what type of WU to send, if any, based on the configuration of the computer requesting the work unit. While it would be more maintenance, perhaps it is within the realm of possibility to keep both and send only 4.2 apps to computers to that support it?

Aside from that, I am highly in favor of making the long queue 4.2 only if limitations exist that require that each queue is one app or the other - should both apps be kept, of course.
____________

Profile GDF
Volunteer moderator
Project administrator
Project developer
Project tester
Volunteer developer
Volunteer tester
Project scientist
Send message
Joined: 14 Mar 07
Posts: 1895
Credit: 629,356
RAC: 0
Level
Gly
Scientific publications
watwatwatwatwat
Message 27500 - Posted: 2 Dec 2012 | 9:26:10 UTC - in response to Message 27499.

In theory the BOINC server looks at the driver and decide what to send you.
Rarely, it sends it the wrong way though.

Even more important to switch to the new app. As I said the new app will also reduce by 1/3 the upload size reducing strain on your broadband and our server.

We might anticipate it to mid december if nobody is against it.

gdf

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 1844
Credit: 10,648,138,594
RAC: 9,965,763
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 27501 - Posted: 2 Dec 2012 | 10:18:03 UTC - in response to Message 27499.
Last modified: 2 Dec 2012 | 10:18:18 UTC

From my experience, 3.1 apps tend to take substantially more time to run than the same type of WU running with the 4.2 app.

Take, for instance, this 3.1 WU - a Nathan respawn http://www.gpugrid.net/workunit.php?wuid=3863083
and this 4.2 WU - also a Nathan respawn http://www.gpugrid.net/workunit.php?wuid=3846597

Note that both were on my 580 machine, granted the same amount of credit, and that the 3.1 WU took almost twice as long as the 4.2 WU. Personally, I find situations like this particularly frustrating since it seems extremely inefficient to run 3.1 apps on machines that support the 4.2 app.

You can use my workaround to fix this problem on your side.

Wiyosaya
Send message
Joined: 22 Nov 09
Posts: 111
Credit: 171,819,453
RAC: 392,567
Level
Ile
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 27622 - Posted: 9 Dec 2012 | 3:50:17 UTC - in response to Message 27501.

From my experience, 3.1 apps tend to take substantially more time to run than the same type of WU running with the 4.2 app.

Take, for instance, this 3.1 WU - a Nathan respawn http://www.gpugrid.net/workunit.php?wuid=3863083
and this 4.2 WU - also a Nathan respawn http://www.gpugrid.net/workunit.php?wuid=3846597

Note that both were on my 580 machine, granted the same amount of credit, and that the 3.1 WU took almost twice as long as the 4.2 WU. Personally, I find situations like this particularly frustrating since it seems extremely inefficient to run 3.1 apps on machines that support the 4.2 app.

You can use my workaround to fix this problem on your side.

Thanks.

I had not checked this thread in a while, however, if I get another 3.1 app on my machine, I will try this. 3.1 apps are especially slow for my 460. Most often, it is unable to crunch a 3.1 WU within 24 hours. However, my 460 is able to complete most 4.2 WUs in something like 12 hours. 3.1 v 4.2 on the 460 makes a substantial difference. It is almost as if I had bought another card. :)
____________

Profile Stoneageman
Avatar
Send message
Joined: 25 May 09
Posts: 211
Credit: 12,252,435,346
RAC: 8,631,734
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 27739 - Posted: 19 Dec 2012 | 20:04:47 UTC - in response to Message 27500.

In theory the BOINC server looks at the driver and decide what to send you.
Rarely, it sends it the wrong way though.

Even more important to switch to the new app. As I said the new app will also reduce by 1/3 the upload size reducing strain on your broadband and our server.

We might anticipate it to mid december if nobody is against it.

gdf


The last few days, 15% of tasks sent to me have been 3.1. Not what I'd call rarely. Any update on bringing the new app into production.

Profile GDF
Volunteer moderator
Project administrator
Project developer
Project tester
Volunteer developer
Volunteer tester
Project scientist
Send message
Joined: 14 Mar 07
Posts: 1895
Credit: 629,356
RAC: 0
Level
Gly
Scientific publications
watwatwatwatwat
Message 27758 - Posted: 20 Dec 2012 | 13:58:08 UTC - in response to Message 27739.

We can remove already 3.1 apps on the long queue, but we will not upgrade the application until new year.

gdf

Profile dskagcommunity
Avatar
Send message
Joined: 28 Apr 11
Posts: 456
Credit: 810,073,458
RAC: 0
Level
Glu
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwat
Message 27765 - Posted: 20 Dec 2012 | 16:31:34 UTC

+1
____________
DSKAG Austria Research Team: http://www.research.dskag.at



Crunching for my deceased Dog who had "good" Braincancer..

Wiyosaya
Send message
Joined: 22 Nov 09
Posts: 111
Credit: 171,819,453
RAC: 392,567
Level
Ile
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 27879 - Posted: 30 Dec 2012 | 5:11:33 UTC
Last modified: 30 Dec 2012 | 5:11:59 UTC

I tried the suggestion above, and it seemed to work once. However, the same machine just got another Cuda31 job, and by estimating the time that it will take from the current progress, it looks like it is again running Cuda31 rather than Cuda42. According to progress vs run time, this will finish in roughly 48 hours on my machine. After applying the work-around, is this supposed to run all subsequent 31 tasks as 42 tasks? Noeilla's usually run in about 23 hours on my 460 with 42 rather than what looks like 48 for this 31 job.
____________

Post to thread

Message boards : News : CUDA upgrade notice