Advanced search

Message boards : Graphics cards (GPUs) : Cuda 4.2 WUs seem much faster than the old 3.1

Author Message
Mark Henderson
Send message
Joined: 21 Dec 08
Posts: 51
Credit: 26,320,167
RAC: 0
Level
Val
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwat
Message 25733 - Posted: 17 Jun 2012 | 4:20:56 UTC
Last modified: 17 Jun 2012 | 5:12:46 UTC

Is it just me or is the cuda 4.2 WUs 30-40 percent faster than the 3.1s are.
It seems amazing the difference is so extreme.

Also, is it a good time now to get a 680 ? now that the project will support it ??

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2343
Credit: 16,201,255,749
RAC: 6,169
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25734 - Posted: 17 Jun 2012 | 8:13:41 UTC - in response to Message 25733.
Last modified: 17 Jun 2012 | 8:14:22 UTC

Is it just me or is the cuda 4.2 WUs 30-40 percent faster than the 3.1s are.

It's much faster on every host.

It seems amazing the difference is so extreme.

It is amazing, but it will vary among different wourkunits.

Also, is it a good time now to get a 680 ? now that the project will support it ??

Definitely!

Profile dskagcommunity
Avatar
Send message
Joined: 28 Apr 11
Posts: 456
Credit: 817,865,789
RAC: 0
Level
Glu
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25735 - Posted: 17 Jun 2012 | 10:11:34 UTC
Last modified: 17 Jun 2012 | 10:13:52 UTC

every host? i still get app 3.1 wus on the 285gtx O.o is it because cc1.3? or must i reattach?

on the 560ti i got 4.2 and i was supprised it done a short WU in 4 hours (before i think it was 5,5h?), so nice speed there ^^
____________
DSKAG Austria Research Team: http://www.research.dskag.at



Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25741 - Posted: 17 Jun 2012 | 12:48:08 UTC - in response to Message 25735.

The improvement is very noticeable on the normal length tasks, and certainly more than the expected 15% improvement. However it does vary considerably, from task to task!
I'll quote myself from here:

    BTW the performance improvement I'm seeing on my CC2.0 Fermi between running tasks on the CUDA4.2 and CUDA3.1 app varies from ~23% to ~56%

    IBUCH_3_affTRYP ~23% faster
    PAOLA_3EKO ~36% faster
    MJHARVEY_MJHXA1 ~56% faster


Different setups will have somewhat different performance improvements; GPU type/spec, CPU type (performance), free CPU count, RAM type... The more the app is optimized the greater the impact of the performance from other resources; the greater the potential for bottlenecks.

I expect similar improvements for the long app when it arrives.
____________
FAQ's

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

Profile Carlesa25
Avatar
Send message
Joined: 13 Nov 10
Posts: 328
Credit: 72,619,453
RAC: 206
Level
Thr
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25742 - Posted: 17 Jun 2012 | 12:48:29 UTC - in response to Message 25735.
Last modified: 17 Jun 2012 | 12:49:28 UTC

Hi: All tasks are downloaded 4.2 CUDA GTX295 mistake on my Ubuntu 12.04-64bit driver 295.59 the CUDA 3.1 are working perfectly.

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25743 - Posted: 17 Jun 2012 | 13:49:01 UTC - in response to Message 25742.
Last modified: 17 Jun 2012 | 13:53:37 UTC

process exited with code 247 (0xf7, -9)

This might result from a missing/incompatible Library file. Check if there are any updates.

Is anyone else having this problem?

I know there are a few people with Kernal/driver mismatches:

process exited with code 255 (0xff, -1)
</message>
<stderr_txt>
Error: API mismatch: the NVIDIA kernel module has version 295.40,
but this NVIDIA driver component has version 295.53. Please make
sure that the kernel module and all NVIDIA driver components
have the same version.

Perhaps there is a general problem with CC1.3 cards and the 4.2app on Linux?
____________
FAQ's

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

ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2705
Credit: 1,311,122,549
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25745 - Posted: 17 Jun 2012 | 14:17:51 UTC - in response to Message 25743.

Or a problem in some Linux driver packages?

MrS
____________
Scanning for our furry friends since Jan 2002

TheFiend
Send message
Joined: 26 Aug 11
Posts: 99
Credit: 2,500,112,138
RAC: 0
Level
Phe
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25749 - Posted: 17 Jun 2012 | 15:43:16 UTC

Switched to cuda4.2 units for today.... so far 2 completed units and 3 have failed exit code - 98 (0x62)

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25751 - Posted: 17 Jun 2012 | 18:05:41 UTC - in response to Message 25749.

If you updated the driver after starting a task, that task will probably fail.
____________
FAQ's

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

TheFiend
Send message
Joined: 26 Aug 11
Posts: 99
Credit: 2,500,112,138
RAC: 0
Level
Phe
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25752 - Posted: 17 Jun 2012 | 21:22:53 UTC - in response to Message 25751.

If you updated the driver after starting a task, that task will probably fail.


Been running the updated driver for a couple of weeks and 3.1 tasks (long) have been running ok.

Now had 6 failures and 2 completed.

My crunchers will be going offline whilst I jet off for a family holiday. Will see how things run when I get back.

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2343
Credit: 16,201,255,749
RAC: 6,169
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25753 - Posted: 17 Jun 2012 | 22:25:54 UTC - in response to Message 25752.
Last modified: 17 Jun 2012 | 22:29:10 UTC

Now had 6 failures and 2 completed.

The CUDA4.2 failures could be caused by overclocking adjusted for CUDA3.1 tasks.
My GTX 480s need 25mV more for crunching CUDA4.2 tasks without failures, also my GTX 590s had to be set to 625MHz instead of 725MHz (I don't want to raise their voltage).

Profile Carlesa25
Avatar
Send message
Joined: 13 Nov 10
Posts: 328
Credit: 72,619,453
RAC: 206
Level
Thr
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25754 - Posted: 17 Jun 2012 | 22:42:25 UTC - in response to Message 25753.
Last modified: 17 Jun 2012 | 22:43:32 UTC

Hi, I reinstalled the Nvidia 295.59 driver without problem, the OS works perfectly and the GTX295 is two tasks normally CUDA 3.1.ACEMD2

We'll see if any task more CUDA 4.2.

TheFiend
Send message
Joined: 26 Aug 11
Posts: 99
Credit: 2,500,112,138
RAC: 0
Level
Phe
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25755 - Posted: 17 Jun 2012 | 23:41:07 UTC - in response to Message 25753.

Now had 6 failures and 2 completed.

The CUDA4.2 failures could be caused by overclocking adjusted for CUDA3.1 tasks.
My GTX 480s need 25mV more for crunching CUDA4.2 tasks without failures, also my GTX 590s had to be set to 625MHz instead of 725MHz (I don't want to raise their voltage).


Will try fiddling around with settings when I return from holiday.

If the long queue is switching to CUDA4.2 I'll try and get the cruncher stable on shorties first.

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25757 - Posted: 18 Jun 2012 | 0:10:57 UTC - in response to Message 25756.

The problems can wait. Enjoy your holiday ;)
____________
FAQ's

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

Profile Carlesa25
Avatar
Send message
Joined: 13 Nov 10
Posts: 328
Credit: 72,619,453
RAC: 206
Level
Thr
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25764 - Posted: 18 Jun 2012 | 18:11:18 UTC - in response to Message 25757.

I still have errors in CUDA 4.2, Nvidia driver 302.17 installed new, not GTX295 OC.

CUDA 3.1 is working perfectly

Error: "process exited with code 247 (0xf7, -9)"

Rantanplan
Send message
Joined: 22 Jul 11
Posts: 166
Credit: 138,629,987
RAC: 0
Level
Cys
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25765 - Posted: 18 Jun 2012 | 19:59:57 UTC
Last modified: 18 Jun 2012 | 20:37:07 UTC

i am here with XP and 301.42, still getting errors with 2x 460 GTX.

Now its some new on cuda 4.2 PAOLA, a bluescreen, tell that:

nv4_disp

0x000000EA
0x89788770
0x89FC06C8
0xB84DBCBC
0x00000001

Could that say someting !?

Edit:

ok, raised the voltage by 25mV, now waiting and watching.

ok 30 minutes , and no error. yeepee !!!

Profile dskagcommunity
Avatar
Send message
Joined: 28 Apr 11
Posts: 456
Credit: 817,865,789
RAC: 0
Level
Glu
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25772 - Posted: 19 Jun 2012 | 15:20:41 UTC

No one answered me for some days now, hope it was only a overreading ^^ Are the cuda 42 apps only for fermi&kepler, because i get only cuda31 apps (yes i tested short WUs) on my 285gtx or must i reattach this machine? (what i must time right because its a unattended machine)
____________
DSKAG Austria Research Team: http://www.research.dskag.at



ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2705
Credit: 1,311,122,549
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25777 - Posted: 19 Jun 2012 | 17:48:53 UTC - in response to Message 25772.

You're running driver: 301.42, so this shouldn't be an issue. Maybe you selected one the queue which doesn't have the new app yet?

MrS
____________
Scanning for our furry friends since Jan 2002

Profile Carlesa25
Avatar
Send message
Joined: 13 Nov 10
Posts: 328
Credit: 72,619,453
RAC: 206
Level
Thr
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25778 - Posted: 19 Jun 2012 | 18:21:53 UTC - in response to Message 25777.

4.2 CUDA tasks keep getting error on my GTX295 Linux.

Can anyone confirm if they only work the FERMI and KEPLER ... is to avoid wasting time, yours and mine. Thank you.

Profile dskagcommunity
Avatar
Send message
Joined: 28 Apr 11
Posts: 456
Credit: 817,865,789
RAC: 0
Level
Glu
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25780 - Posted: 19 Jun 2012 | 20:12:58 UTC
Last modified: 19 Jun 2012 | 20:14:56 UTC

I tried both long and short. So this shouldnt the problem. But ok it seems there are still some probs with cuda 42 apps when i read the threads and i need valid results not wasting time :) only wanted to try one if it works or not as info for the team.
____________
DSKAG Austria Research Team: http://www.research.dskag.at



Paul Raney
Send message
Joined: 26 Dec 10
Posts: 115
Credit: 416,576,946
RAC: 0
Level
Gln
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwat
Message 25801 - Posted: 21 Jun 2012 | 11:45:41 UTC - in response to Message 25780.

Should we migrate our Fermi cards to Cuda 4.2 compliant drivers? What drivers are required? Do we have a new recommendation on the driver version to be used?

The 301.42 drivers work well on one of my systems. Is it recommended that I upgrade all of my systems?

____________
Thx - Paul

Note: Please don't use driver version 295 or 296! Recommended versions are 266 - 285.

Profile MJH
Project administrator
Project developer
Project scientist
Send message
Joined: 12 Nov 07
Posts: 696
Credit: 27,266,655
RAC: 0
Level
Val
Scientific publications
watwat
Message 25802 - Posted: 21 Jun 2012 | 12:57:20 UTC - in response to Message 25801.

Should we migrate our Fermi cards to Cuda 4.2 compliant drivers? What drivers are required? Do we have a new recommendation on the driver version to be used?


Yes, if you can. The 4.2 app is fast. The latest drivers (301 series) are fine.

Matt


klepel
Send message
Joined: 23 Dec 09
Posts: 189
Credit: 4,196,536,293
RAC: 1,518,485
Level
Arg
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25803 - Posted: 21 Jun 2012 | 13:24:19 UTC

No, I had problems with my GTX570 SC card. More from the office.

ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2705
Credit: 1,311,122,549
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25808 - Posted: 21 Jun 2012 | 19:34:04 UTC - in response to Message 25804.

Does the card feature a substantial factory-OC and is failing some tasks with 4.2? It has been observed that the new app is somewhat less tolerant to overclocking, or demands higherstability - however you want to look at it.

MrS
____________
Scanning for our furry friends since Jan 2002

5pot
Send message
Joined: 8 Mar 12
Posts: 411
Credit: 2,083,882,218
RAC: 0
Level
Phe
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25812 - Posted: 21 Jun 2012 | 21:51:43 UTC

MJH:

No doubt it's fast. First time I've ever seen that short queue get depleted so quickly ;)

Seems like the hopper needs more added in quite often now...


As a side note, any idea one when we can finally start taking a crack at those long queue tasks. Seems like, from what I read in the forums, that everything seems to being a-ok!!

Cheers

PS, Yours seem to have the greatest speed difference gain. Congrats!!

klepel
Send message
Joined: 23 Dec 09
Posts: 189
Credit: 4,196,536,293
RAC: 1,518,485
Level
Arg
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25813 - Posted: 21 Jun 2012 | 23:17:02 UTC - in response to Message 25803.

No, I had problems with my GTX570 SC card. More from the office.


Sorry I have not written during the day.

I transferred my GTX 570 SC to another computer to make room for my GTX670 SC in my most potent PC. At the same time I did up-date drivers on both computers to 301.42. All short work units Cuda 42 work well on both systems. However when I have attached the computer with the GTX 570 SC card to the long work queue, this long work unites (Cuda 31) failed. Means the computer shuts down randomly. I was able to restart computer as well as continuo the work unites, but the computer shuts down again after half an hour or so.

The GTX 570 SC card worked very well short and long work units with Driver 275.33 on the computer, which has actually the GTX 670.

The GTX 570 SC card works actually well with Primegrid Genefer (cuda32_13) and PPS (Sieve) (cuda23), as well with Einstein (BRP4cuda32nv301), so it is not the Cuda version.

The video cards are factory overclocked. I have only modified the fan speed, so the temperature of the cards is below 70º C.

5pot
Send message
Joined: 8 Mar 12
Posts: 411
Credit: 2,083,882,218
RAC: 0
Level
Phe
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25814 - Posted: 22 Jun 2012 | 0:29:38 UTC

I only see two workunits which have failed with energies have become nan. Maybe increase the voltage slightly and see if this helps stability? Or of energy usage is a concern you could try underclocking it as well.

Paul Raney
Send message
Joined: 26 Dec 10
Posts: 115
Credit: 416,576,946
RAC: 0
Level
Gln
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwat
Message 25817 - Posted: 22 Jun 2012 | 3:31:18 UTC - in response to Message 25815.

I recently had some issues with work units causing the computer to hang or freeze. The only way to clear the condition was a reset of the computer (or power cycle). It happened a couple of times to one of my systems that never has issues. I did not change drivers or any of the OC settings on the system. I assumed it was a couple of bad work units.

Try the 570 SC with the new drivers again. I bet you will find it was bad work units and nothing to do with the driver.

Best of Luck!
____________
Thx - Paul

Note: Please don't use driver version 295 or 296! Recommended versions are 266 - 285.

Profile dskagcommunity
Avatar
Send message
Joined: 28 Apr 11
Posts: 456
Credit: 817,865,789
RAC: 0
Level
Glu
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25819 - Posted: 22 Jun 2012 | 11:00:52 UTC

Paul i had/have that too with some nathans rps ( perhaps it was nathan too in your case??) at the time, some units freeze the machine in a special way. Desktop with no icons but windows is running. Often @ the beginning of a new computation while he uploads the finished wu. Good to hear they are the wus perhaps but itcomputes to finishing line after reboot. Strange. Changed fan to manual to bring it from 90 to 80 degrees. Will see how it works over weekend bacause this is litle annoing on a unaatended machine.
____________
DSKAG Austria Research Team: http://www.research.dskag.at



MarkJ
Volunteer moderator
Volunteer tester
Send message
Joined: 24 Dec 08
Posts: 738
Credit: 200,909,904
RAC: 0
Level
Leu
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25820 - Posted: 22 Jun 2012 | 13:10:52 UTC - in response to Message 25817.
Last modified: 22 Jun 2012 | 13:20:06 UTC

I recently had some issues with work units causing the computer to hang or freeze. The only way to clear the condition was a reset of the computer (or power cycle). It happened a couple of times to one of my systems that never has issues. I did not change drivers or any of the OC settings on the system. I assumed it was a couple of bad work units.

Try the 570 SC with the new drivers again. I bet you will find it was bad work units and nothing to do with the driver.

Best of Luck!


Probably haven't made the apps threadsafe which is why you need to reboot the machine. I know Richard asked the same question back in March but never got a reply. See this message

Edit:
Found the original message with code snippets here

Maybe GDF or one of the other developers can confirm Acemd 6.16 (cuda42) does use the above logic (ie its threadsafe)
____________
BOINC blog

Paul Raney
Send message
Joined: 26 Dec 10
Posts: 115
Credit: 416,576,946
RAC: 0
Level
Gln
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwat
Message 25821 - Posted: 22 Jun 2012 | 13:11:46 UTC - in response to Message 25819.

Wow - 90 is hot! I flashed my EVGA GTX cards with the newest BIOS and it unlocks the fans to 100%. I have them set at 90% at 70C and my 580s are down to 72C - 74C

This summer is going to be a challenge to keep these systems cool.

Dry Ice could be a solution!

BTW - is dry ice safe for the computers?
____________
Thx - Paul

Note: Please don't use driver version 295 or 296! Recommended versions are 266 - 285.

klepel
Send message
Joined: 23 Dec 09
Posts: 189
Credit: 4,196,536,293
RAC: 1,518,485
Level
Arg
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25823 - Posted: 22 Jun 2012 | 15:38:08 UTC - in response to Message 25814.

I only see two workunits which have failed with energies have become nan. Maybe increase the voltage slightly and see if this helps stability? Or of energy usage is a concern you could try underclocking it as well.

As I have said, the problems are only with the cuda 31 app, not with cuda 42. In the night the computer failed with two cuda 31 apps. The strange thing: When I restarted the computer again, blue screen with warning message and the GPUGRID project was detached afterwards, so I attached anew and got one new work unit with cuda 42. Hope this works out just fine. The other two are just lost, and won´t be finished ever.

Side note: I am running malariacontrol.net at all computers at the moment and this writes heavily on hard disks, so this might also cause the observed problems. Also my co-workers experience problems with this project, whereas with climateprediction,net - my normal CPU load - I normally do not have problems.

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25826 - Posted: 22 Jun 2012 | 15:55:09 UTC - in response to Message 25823.
Last modified: 22 Jun 2012 | 16:02:09 UTC

The CPU tasks may be the source of some of your problems.

On your 6core all the failures were 'Aborted by user'.
On your GTX670 system, one nan failure might have been as a result of the CPU projects, the other failure happened because you started running the task on a GTX570 and then continued to run it on a GTX670.
The 9800GT error was the old cufft issue, common with CC1.1 cards.

How many CPU cores are you using to crunch CPU tasks? I know you have a 6core, an 8core and an 8thread system.
Generally it's more stable to use one less CPU core/thread than the total. This also tends to expedite the GPU tasks and even the other CPU tasks slightly, so you gain stability and GPU task performance at a very limited CPU expense.
It's better to run a mix of CPU task types, for lots of reasons.
It's advisable to increase you write to disk times on anything more than a dual core. I use 600 to 900sec.
It's also beneficial to use a second drive for the Boinc data directory, where possible.
____________
FAQ's

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

klepel
Send message
Joined: 23 Dec 09
Posts: 189
Credit: 4,196,536,293
RAC: 1,518,485
Level
Arg
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25828 - Posted: 22 Jun 2012 | 16:25:01 UTC - in response to Message 25826.

Yes it might be that the CPU task cause the problems. Unfortunately my computer knowledge is very limited, so what goes behind Boinc Manager, will have to be explained step by step. So I will be able to try it on my own, which I would be willing to do.

The GTX670 on the 8-core does not worry me that much, because it is true, I changed the GPU, and the other might happen ones in a while.

The 9800GT will not crunch GPUGRID anymore as intend to use only GTX570 and GTX670 for this project, lower end cards I will use for Primegrid, as these do not work very well with GPUGRID.

Which really worries me is the GTX570 on the 6core computer.

All CPU cores are crunching CPU projects, I have seen that you disable one core for the GPU with Svan_sync (I think), which I am willing to do but I do not see how. However with climateprediction.net (my main focus) I never had any issues with using GPUs, as well as write to disk times.

That said I normally running climteprediction.net on all computers, malariacontrol.net is only my “back-up” project, if the first runs out of work.

So if you are willing to guide me, you are welcome!

8core, has solid state disk and normal hard disk – but if I remember correctly I am writing boinc to solid state disk as I assumed this will improve speed. 6core has a small solid state disk for the programs as all data is loaded to an external disk.

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25829 - Posted: 22 Jun 2012 | 16:29:46 UTC - in response to Message 25828.
Last modified: 22 Jun 2012 | 16:36:30 UTC

When you reattached to the project you lost any GPU tasks that were running. These must just appear as 'Aborted by user'.
Did you reattach twice?

I suggest you configure Boinc to use one less CPU core on your 6core system, and write to disk less frequently:
Boinc Manager (Advanced View),
Tools Tab,
Computing Preferences,
Processor Usage Tab,
On multiprocessor systems use at most 99% of the processors,
Disk and Memory Usage Tab,
Tasks Checkpoint to Disk at most every 900 seconds,
OK.
____________
FAQ's

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

klepel
Send message
Joined: 23 Dec 09
Posts: 189
Credit: 4,196,536,293
RAC: 1,518,485
Level
Arg
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25832 - Posted: 22 Jun 2012 | 17:08:36 UTC - in response to Message 25829.

When you reattached to the project you lost any GPU tasks that were running. These must just appear as 'Aborted by user'.
Did you reattach twice?


No, I reattached once.
It is: Task 5522912 and Task 5522851. Both are still marked in Progress on GPUGRID Homepage. But are not mentioned any more in Boinc Manager.

KING100N
Send message
Joined: 27 Mar 09
Posts: 2
Credit: 113,212,081
RAC: 0
Level
Cys
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwat
Message 25844 - Posted: 24 Jun 2012 | 8:28:19 UTC

Should i have something besides boinc 7.0.25 and 301.42 drivers to get cuda 4.2 tasks

Snow Crash
Send message
Joined: 4 Apr 09
Posts: 450
Credit: 539,316,349
RAC: 0
Level
Lys
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25845 - Posted: 24 Jun 2012 | 9:35:00 UTC - in response to Message 25844.

Currently the CUDA 4.2 app is only being used for the Standard WUs and it looks like you are only processing the Longs. Hopefully GDF will release it next week? for the Longs also.
____________
Thanks - Steve

Profile dskagcommunity
Avatar
Send message
Joined: 28 Apr 11
Posts: 456
Credit: 817,865,789
RAC: 0
Level
Glu
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25846 - Posted: 24 Jun 2012 | 16:29:33 UTC - in response to Message 25844.
Last modified: 24 Jun 2012 | 16:30:12 UTC

Should i have something besides boinc 7.0.25 and 301.42 drivers to get cuda 4.2 tasks


I have the same problem ^^ Hope somebody tells us if we need to reattach or something ^^ (an nooo i tried short/standart units and it worked on a fermi card, only not on the 285gtx)

I must wait until a WU is finished AND im in the room where the machine is, then i will try a reattach, could need a few days :)
____________
DSKAG Austria Research Team: http://www.research.dskag.at



Profile [PUGLIA] kidkidkid3
Avatar
Send message
Joined: 23 Feb 11
Posts: 79
Credit: 953,228,044
RAC: 1,959,858
Level
Glu
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25849 - Posted: 25 Jun 2012 | 10:27:15 UTC

Hi all,
my pc is running first longqueue WU with Cuda 4.2 (a Paola RND).
The expected time on GTS450 (with GF116) under windows xp will be 17 hours !
Here the result of Gpu-z
url]http://gpuz.techpowerup.com/12/06/25/3aa.png[/url]
Good work, great !
K.
____________
Dreams do not always come true. But not because they are too big or impossible. Why did we stop believing.
(Martin Luther King)

MarkJ
Volunteer moderator
Volunteer tester
Send message
Joined: 24 Dec 08
Posts: 738
Credit: 200,909,904
RAC: 0
Level
Leu
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25850 - Posted: 25 Jun 2012 | 10:34:53 UTC - in response to Message 25846.

Should i have something besides boinc 7.0.25 and 301.42 drivers to get cuda 4.2 tasks


I have the same problem ^^ Hope somebody tells us if we need to reattach or something ^^ (an nooo i tried short/standart units and it worked on a fermi card, only not on the 285gtx)

I must wait until a WU is finished AND im in the room where the machine is, then i will try a reattach, could need a few days :)


Detach/reattach wouldn't help. My understanding is they are only giving cuda42 to compute capability 2 or higher cards. The gtx285 is only a compute capability of 1.3
____________
BOINC blog

Snow Crash
Send message
Joined: 4 Apr 09
Posts: 450
Credit: 539,316,349
RAC: 0
Level
Lys
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25852 - Posted: 25 Jun 2012 | 10:48:32 UTC
Last modified: 25 Jun 2012 | 11:27:36 UTC

WinXP32, BOINC 7.0.25, GTX480, GTX295 301.42

I am running the 4.2 app on my 480 and 295 (same box) and the MJHARVEY WUs are at 99% utilization!

I don't have accurate runtime estimates as the 480 is 2x as fast as the 295 so my numbers are all over the place :-) but I will do a comparison to the 3.1 WUs once I have completed some of the 4.2 WUs.

No re-attach was necessary.


Win7 x64, BOINC 7.0.25, GTX680 301.42
Now running an 78-IBUCH_mutTRYP at 95% utilization WOOT. Uses a full CPU core but I'm good with that :-)
____________
Thanks - Steve

Profile Carlesa25
Avatar
Send message
Joined: 13 Nov 10
Posts: 328
Credit: 72,619,453
RAC: 206
Level
Thr
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25858 - Posted: 25 Jun 2012 | 14:01:02 UTC
Last modified: 25 Jun 2012 | 14:01:54 UTC

Hi, All Tasks CUDA 4.2 give me error on my Ubuntu 12.04 GTX 295 driver 302.17.

The tasks are CUDA 4.2 only for KEPLER or FERMI...?

Snow Crash
Send message
Joined: 4 Apr 09
Posts: 450
Credit: 539,316,349
RAC: 0
Level
Lys
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25861 - Posted: 25 Jun 2012 | 15:33:19 UTC
Last modified: 25 Jun 2012 | 15:35:00 UTC

My first completed 4.2 on GTX680 is a NATHAN_RPS on Win7 x64 = 13,945.16!
____________
Thanks - Steve

5pot
Send message
Joined: 8 Mar 12
Posts: 411
Credit: 2,083,882,218
RAC: 0
Level
Phe
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25862 - Posted: 25 Jun 2012 | 16:04:00 UTC
Last modified: 25 Jun 2012 | 16:12:15 UTC

670 rig:

MJH long run 5%@ 15 minutes.

EDIT: using 92-94 GPU (non-HT rig) with a 67% Power draw according to Precision.

With a 170W spec from NVIDIA, this puts the card at 114W (not read from wall)

EDIT 2x:

Interestingly, my 680 is running a MJH long as well, and is getting about the same utilization (HT rig) but with 87% draw?

Profile dskagcommunity
Avatar
Send message
Joined: 28 Apr 11
Posts: 456
Credit: 817,865,789
RAC: 0
Level
Glu
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25883 - Posted: 26 Jun 2012 | 6:49:40 UTC

Oh wow i got my first cuda 42 app on my 285gtx without reattach. :) will see how good it works.
____________
DSKAG Austria Research Team: http://www.research.dskag.at



shdbcamping
Send message
Joined: 2 May 12
Posts: 22
Credit: 145,756,579
RAC: 0
Level
Cys
Scientific publications
watwatwatwatwatwatwat
Message 25938 - Posted: 27 Jun 2012 | 18:47:54 UTC

Eight CUDA 4.2 long apps completed. Zero errors so far on my gtx570's. Seem to be 30-40% faster than like pointed CUDA 3.1 work done previously. I made no changes since updating drivers to 301.42. They just started showing up.
So far I gotta say that I'm loving them!

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2343
Credit: 16,201,255,749
RAC: 6,169
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25940 - Posted: 27 Jun 2012 | 19:00:40 UTC - in response to Message 25938.
Last modified: 27 Jun 2012 | 19:01:10 UTC

On my laptop's GT 630M a CUDA4.2 NATHAN_RPS ran 93% faster than the CUDA3.1 (122252 sec vs 236314 sec). This way it could even catch the 48h bonus. :) That's really amazing! Could it be that the lesser cards gain more than the high end ones?

Paul Raney
Send message
Joined: 26 Dec 10
Posts: 115
Credit: 416,576,946
RAC: 0
Level
Gln
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwat
Message 25944 - Posted: 27 Jun 2012 | 22:22:55 UTC - in response to Message 25940.

My 570s appear to gain more than my 580s so CUDA 4.2 might help the lesser cards the most.
____________
Thx - Paul

Note: Please don't use driver version 295 or 296! Recommended versions are 266 - 285.

shdbcamping
Send message
Joined: 2 May 12
Posts: 22
Credit: 145,756,579
RAC: 0
Level
Cys
Scientific publications
watwatwatwatwatwatwat
Message 25950 - Posted: 28 Jun 2012 | 1:49:11 UTC

Any news on the shorter openCL WU's? Have they become stable and/or available?

ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2705
Credit: 1,311,122,549
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25979 - Posted: 28 Jun 2012 | 18:14:25 UTC - in response to Message 25950.

OpenCL? This is all CUDA.

MrS
____________
Scanning for our furry friends since Jan 2002

Profile vitalidze
Send message
Joined: 15 May 09
Posts: 20
Credit: 239,712,351
RAC: 0
Level
Leu
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 25986 - Posted: 29 Jun 2012 | 6:11:32 UTC

PPD really rocks!!! I think within a couple of days we will have much more defectors from primegrid or something like this. Strengthen your servers)

shdbcamping
Send message
Joined: 2 May 12
Posts: 22
Credit: 145,756,579
RAC: 0
Level
Cys
Scientific publications
watwatwatwatwatwatwat
Message 26024 - Posted: 30 Jun 2012 | 4:15:39 UTC - in response to Message 25979.

OpenCL? This is all CUDA.

MrS

Several versions being assigned in both Long and short. Cuda (non openCL) is 3.1 and lower. It will show in boinc manager. OpenCL 4.2 work unit is not the same.
CUDA all true. Programming different, I would suppose.
:)

TheFiend
Send message
Joined: 26 Aug 11
Posts: 99
Credit: 2,500,112,138
RAC: 0
Level
Phe
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 26028 - Posted: 30 Jun 2012 | 8:25:22 UTC - in response to Message 25752.



Been running the updated driver for a couple of weeks and 3.1 tasks (long) have been running ok.

Now had 6 failures and 2 completed.

My crunchers will be going offline whilst I jet off for a family holiday. Will see how things run when I get back.


Back off me hols :)

Returned both my cards to stock speeds and ran 2 short WU's as a test and both completed. Now running long WU's and both seem quite happy.

When we were in Thailand I was nearly tempted to buy a GTX560 to add to my collection - they were quite a bit cheaper out there than in the UK.

ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2705
Credit: 1,311,122,549
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 26037 - Posted: 30 Jun 2012 | 11:42:19 UTC - in response to Message 26024.

OpenCL 4.2 work unit is not the same.

4.2 is the CUDA version. This has nothing to do with OpenCL.

MrS
____________
Scanning for our furry friends since Jan 2002

Post to thread

Message boards : Graphics cards (GPUs) : Cuda 4.2 WUs seem much faster than the old 3.1

//