Running SETI@home on an nVidia Fermi GPU

Message boards : Number crunching : Running SETI@home on an nVidia Fermi GPU
Message board moderation

To post messages, you must log in.

Previous · 1 · 2 · 3 · 4 · 5 · 6 . . . 15 · Next

AuthorMessage
Profile MadMaC
Volunteer tester
Avatar

Send message
Joined: 4 Apr 01
Posts: 201
Credit: 47,158,217
RAC: 0
United Kingdom
Message 1002062 - Posted: 9 Jun 2010, 8:06:46 UTC

Think I screwed this up last night, but will have another look when I get back tomorrow. I think it was adding the app block for the rescheduler entries that mess things up, so I rolled back to my original entry..

Looking at my tasks online I can see lots of 'Not in DB' messages in the application column, is this due to the scheduler problems posted here and elsewhere?
ID: 1002062 · Report as offensive
TheFreshPrince a.k.a. BlueTooth76
Avatar

Send message
Joined: 4 Jun 99
Posts: 210
Credit: 10,315,944
RAC: 0
Netherlands
Message 1002201 - Posted: 9 Jun 2010, 19:34:30 UTC
Last modified: 9 Jun 2010, 19:35:46 UTC

Bought a GTX 470 today, I hope it all works when the servers start splitting again :)

It's fun to see many old members over here :)
ID: 1002201 · Report as offensive
Richard Haselgrove Project Donor
Volunteer tester

Send message
Joined: 4 Jul 99
Posts: 14679
Credit: 200,643,578
RAC: 874
United Kingdom
Message 1002237 - Posted: 9 Jun 2010, 21:19:53 UTC - in response to Message 1002201.  

I moved my GTX 470 to a different host while we were down yesterday, and left it on stock settings (no app_info) for the time being to see what happens.

Just got an allocation of v6.10, with the right application, DLLs and all. Looks like they've got it sorted, and the Fermis should work as designed without babysitting.
ID: 1002237 · Report as offensive
Profile perryjay
Volunteer tester
Avatar

Send message
Joined: 20 Aug 02
Posts: 3377
Credit: 20,676,751
RAC: 0
United States
Message 1002246 - Posted: 9 Jun 2010, 21:54:13 UTC - in response to Message 1002237.  

Great to hear Richard. Let us know if there are any problems.


PROUD MEMBER OF Team Starfire World BOINC
ID: 1002246 · Report as offensive
Profile SciManStev Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 20 Jun 99
Posts: 6658
Credit: 121,090,076
RAC: 0
United States
Message 1002247 - Posted: 9 Jun 2010, 21:57:47 UTC - in response to Message 1002237.  

I moved my GTX 470 to a different host while we were down yesterday, and left it on stock settings (no app_info) for the time being to see what happens.

Just got an allocation of v6.10, with the right application, DLLs and all. Looks like they've got it sorted, and the Fermis should work as designed without babysitting.


Are you able to get work for your 470 currently? My CPU has no problem getting new work, but the GPU's can't get a thing.

Steve
Warning, addicted to SETI crunching!
Crunching as a member of GPU Users Group.
GPUUG Website
ID: 1002247 · Report as offensive
Richard Haselgrove Project Donor
Volunteer tester

Send message
Joined: 4 Jul 99
Posts: 14679
Credit: 200,643,578
RAC: 874
United Kingdom
Message 1002260 - Posted: 9 Jun 2010, 22:26:48 UTC

It's fetching OK. I'll be off to bed shortly, but you can keep an eye on host 4292666 for me.

Should start reporting shortly after 00:15 UTC (got a GPUGrid to polish off first), run about 8 tasks, then break for three hours or so for another GPUGrid.
ID: 1002260 · Report as offensive
ToxicTBag

Send message
Joined: 5 Feb 10
Posts: 101
Credit: 57,197,902
RAC: 0
United Kingdom
Message 1002646 - Posted: 10 Jun 2010, 19:49:29 UTC

Hi all
I am having the same problem Steve, i installed 6.10.56 and keep getting "reached daily quota of 100tasks" message. Not sure how long this will last, hopefully it is only temporary but it has certainly made me think twice about investing in more hardware.
ID: 1002646 · Report as offensive
Profile SciManStev Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 20 Jun 99
Posts: 6658
Credit: 121,090,076
RAC: 0
United States
Message 1002648 - Posted: 10 Jun 2010, 19:55:58 UTC - in response to Message 1002646.  

Hi all
I am having the same problem Steve, i installed 6.10.56 and keep getting "reached daily quota of 100tasks" message. Not sure how long this will last, hopefully it is only temporary but it has certainly made me think twice about investing in more hardware.


It's not just Fermi's that are having the problem, but many GPU crunchers. You may have a separate problem as I might. See this thread.

http://setiathome.berkeley.edu/forum_thread.php?id=60285&nowrap=true#1002625

We may both have this problem. Todd might also.

Steve
Warning, addicted to SETI crunching!
Crunching as a member of GPU Users Group.
GPUUG Website
ID: 1002648 · Report as offensive
Profile perryjay
Volunteer tester
Avatar

Send message
Joined: 20 Aug 02
Posts: 3377
Credit: 20,676,751
RAC: 0
United States
Message 1002724 - Posted: 10 Jun 2010, 23:11:38 UTC - in response to Message 1002648.  

Might want to see this.. http://setiathome.berkeley.edu/forum_thread.php?id=60287&nowrap=true#1002666 another headache for the Berkeley boys!


PROUD MEMBER OF Team Starfire World BOINC
ID: 1002724 · Report as offensive
Profile Todd Hebert
Volunteer tester
Avatar

Send message
Joined: 16 Jun 00
Posts: 648
Credit: 228,292,957
RAC: 0
United States
Message 1002842 - Posted: 11 Jun 2010, 6:49:52 UTC - in response to Message 1002648.  

No different here - I thought I would wait a day to chime in (and I have been working like a dog) Most of my crunching machines have run out of GPU work and are starving right now. Thankfully my Fermi machine has enough work to keep going through the weekend.

Todd

Hi all
I am having the same problem Steve, i installed 6.10.56 and keep getting "reached daily quota of 100tasks" message. Not sure how long this will last, hopefully it is only temporary but it has certainly made me think twice about investing in more hardware.


It's not just Fermi's that are having the problem, but many GPU crunchers. You may have a separate problem as I might. See this thread.

http://setiathome.berkeley.edu/forum_thread.php?id=60285&nowrap=true#1002625

We may both have this problem. Todd might also.

Steve


ID: 1002842 · Report as offensive
Profile MadMaC
Volunteer tester
Avatar

Send message
Joined: 4 Apr 01
Posts: 201
Credit: 47,158,217
RAC: 0
United Kingdom
Message 1002861 - Posted: 11 Jun 2010, 7:58:18 UTC

Im not too bad, I do have work, but with a 10 day cache, I estimate that I have 20 days work for my cpu and maybe 2 days work for my fermi gpu.
Rescheduler doesn't work for fermi, and I borked a load of units playing with my app info to try and get rescheduler to work, might have another go soon, as my cpu has more work than it can complete within the deadlines and it would make sense to move some of it to the gpu..
ID: 1002861 · Report as offensive
TheFreshPrince a.k.a. BlueTooth76
Avatar

Send message
Joined: 4 Jun 99
Posts: 210
Credit: 10,315,944
RAC: 0
Netherlands
Message 1002959 - Posted: 11 Jun 2010, 14:53:08 UTC

Yes, finally I got it to work :)

<app_info>
<app>
<name>setiathome_enhanced</name>
</app>
<file_info>
<name>AK_v8b_win_x64_SSSE3x.exe</name>
<executable/>
</file_info>
<app_version>
<app_name>setiathome_enhanced</app_name>
<version_num>603</version_num>
<platform>windows_intelx86</platform>
<file_ref>
<file_name>AK_v8b_win_x64_SSSE3x.exe</file_name>
<main_program/>
</file_ref>
</app_version>
<app_version>
<app_name>setiathome_enhanced</app_name>
<version_num>603</version_num>
<platform>windows_x86_64</platform>
<file_ref>
<file_name>AK_v8b_win_x64_SSSE3x.exe</file_name>
<main_program/>
</file_ref>
</app_version>
<app>
<name>setiathome_enhanced</name>
</app>
<file_info>
<name>libfftw3f-3-1-1a_upx.dll</name>
<executable/>
</file_info>
<file_info>
<name>setiathome_6.10_windows_intelx86__cuda_fermi.exe</name>
<executable/>
</file_info>
<file_info>
<name>cudart32_30_14.dll</name>
<executable/>
</file_info>
<file_info>
<name>cufft32_30_14.dll</name>
<executable/>
</file_info>
<app_version>
<app_name>setiathome_enhanced</app_name>
<version_num>610</version_num>
<avg_ncpus>0.200000</avg_ncpus>
<max_ncpus>0.200000</max_ncpus>
<plan_class>cuda_fermi</plan_class>
<file_ref>
<file_name>setiathome_6.10_windows_intelx86__cuda_fermi.exe</file_name>
<main_program/>
</file_ref>
<file_ref>
<file_name>cudart32_30_14.dll</file_name>
</file_ref>
<file_ref>
<file_name>cufft32_30_14.dll</file_name>
</file_ref>
<file_ref>
<file_name>libfftw3f-3-1-1a_upx.dll</file_name>
</file_ref>
<coproc>
<type>CUDA</type>
<count>0.5</count>
</coproc>
</app_version>
</app_info>

My specs:
Windows 7 x64 (Dutch)
Intel E8400 C2D 3.0@3.6Ghz
Asus ENGTX470
2 tasks running on the CPU
2 tasks running on the GTX470 CUDA device
ID: 1002959 · Report as offensive
Richard Haselgrove Project Donor
Volunteer tester

Send message
Joined: 4 Jul 99
Posts: 14679
Credit: 200,643,578
RAC: 874
United Kingdom
Message 1002997 - Posted: 11 Jun 2010, 15:23:42 UTC

Flops estimating for Fermi

In the middle of all the error messages last night, somebody asked about getting the running time estimates for the various applications into balance. This was my initial background explanation:

BOINC 'knows' how fast your CPU and your GPU are. But when you put in an optimised app (strictly: when you use app_info.xml to specify an 'anonymous platform'), you're saying "forget everything you thought you knew: I know better, and I'm telling you everything you need to know in this file."

Except at this stage, you probably haven't put in the speed estimates.

BOINC has to have some idea of speed to work out those 'To completion' times. So it cheats, and looks up the speed of the CPU, and uses that.

But then it goes on to assume that your GPU is the same speed as the CPU - which it patently isn't (well, I hope not, anyway). So your GPU tasks finish sooner than expected, and BOINC gradually adjusts to that - until a CPU task finishes more slowly than BOINC now expects, and we're back to square one.

To cure that habit, you have to put in reasonable figures yourself, which take account not only of the raw speed of the hardware, but also the relative efficiency of the applications you're using.

There's a rough how-to available in app_info for AP503, AP505, MB603 and MB608 - that's the revised version, for afficionados. But it hasn't been brought up-to-date for Fermis.

I'm not going to go over all the practicalities again - read MarkJ's thread, especially the opening post - but here are my experiences with a Fermi.

I was working on my Q9300 host (processor speed 2.5 GHz, benchmark 2.645 GFlops), running BOINC as a service under Windows XP, 32-bit - there are reports that Fermis run significantly slower under Windows 7. My Fermi is a GTX 470, which BOINC rates at 1,089 GFlops peak.

Note that the last figure is peak GFlops. When Mark wrote that earlier thread, BOINC reported NVidia speeds in estimated GFlops, which are 5.6 times lower. My card would rate about 195 GFlops on the old measurement scale.

When I calculate these figures, I aim for a Duration Correction Factor of 0.2. There's nothing special about that figure, except I believe it's roughly what the stock CPU application for SETI settles at on a modern Intel processor (it's a very long time since I tested that assumption!). So if somebody using the stock app stumbles across these message boards, and decides to do all the upgrading at once, their cache shouldn't be too badly affected.

My figures ended up at:

CPU:	    3,826,762,478
GPU:	   57,462,450,464

or, in multiplier terms,

CPU: multiply <p_fpops> by 1.45
GPU: divide peak GFlops by 18.95

That's a rather narrower spread than we're used to with the earlier cards, which I think is perhaps accounted for by:

a) The stock Fermi app uses rather more CPU support than earlier CUDA apps, which drags down CPU performance a bit.
b) The CUDA 3.0 runtime doesn't yet get as close to the theoretical hardware speed as the older DLLs do for their hardware.

Hopefully, both of these will improve as the software matures, but this was a good enough starting point for now.
ID: 1002997 · Report as offensive
Profile perryjay
Volunteer tester
Avatar

Send message
Joined: 20 Aug 02
Posts: 3377
Credit: 20,676,751
RAC: 0
United States
Message 1003019 - Posted: 11 Jun 2010, 15:38:19 UTC - in response to Message 1002997.  

Thanks Richard,

Can you add this to MarkJ's thread so we can have just one thread to point to if we give someone advise on this?


PROUD MEMBER OF Team Starfire World BOINC
ID: 1003019 · Report as offensive
Richard Haselgrove Project Donor
Volunteer tester

Send message
Joined: 4 Jul 99
Posts: 14679
Credit: 200,643,578
RAC: 874
United Kingdom
Message 1003029 - Posted: 11 Jun 2010, 15:51:44 UTC - in response to Message 1003019.  

Thanks Richard,

Can you add this to MarkJ's thread so we can have just one thread to point to if we give someone advise on this?

No - that thread hasn't been added to for over six months, so it's automatically locked. It's also buried away on page 16 (the next thread started will bump it to 17), so I don't think it would be helpful to ask a Mod to resurrect it.

There is also a copy of Mark's opening post in the sticky FAQ's, Astropulse, Cuda, MultiBeam - Read Only, but as the name says that's read only and I can't post there. Perhaps if Pappa is passing he might make a copy.

One reason I posted in this thread is that it is already sticky, and a reasonable gathering of Fermi-related stuff. Would anyone who's posted here mind if I ask a Mod to move some of the individual Q & A material to a separate discussion thread, so we improve the signal-to-noise ratio for future readers? ...

... that is, if I can find a technically-minded Mod: they seem to be rather thin on the ground these days, since OzzFan and the others left us.
ID: 1003029 · Report as offensive
Profile perryjay
Volunteer tester
Avatar

Send message
Joined: 20 Aug 02
Posts: 3377
Credit: 20,676,751
RAC: 0
United States
Message 1003043 - Posted: 11 Jun 2010, 15:56:40 UTC - in response to Message 1003029.  

That's the thread I was thinking about. Didn't remember it was read only but if we could get your information copied over to there it would sure help.


PROUD MEMBER OF Team Starfire World BOINC
ID: 1003043 · Report as offensive
Profile RottenMutt
Avatar

Send message
Joined: 15 Mar 01
Posts: 1011
Credit: 230,314,058
RAC: 0
United States
Message 1003074 - Posted: 11 Jun 2010, 17:32:27 UTC - in response to Message 1003029.  

...they seem to be rather thin on the ground these days, since OzzFan and the others left us.


so that is why it's been so peaceful around here. finally some good news:)

ID: 1003074 · Report as offensive
John

Send message
Joined: 21 May 99
Posts: 51
Credit: 5,667,907
RAC: 0
United States
Message 1003254 - Posted: 11 Jun 2010, 21:24:38 UTC

started the first set of fermi (gxt 470 card)
Hostid http://setiathome.berkeley.edu/show_host_detail.php?hostid=5012909

uninstalled Lunatic Unfied installer
I detached from project first to clear seti@home project file
no app_info file
reattached to project

Fermi files appear to be working no -9 errors yet


ID: 1003254 · Report as offensive
Profile MadMaC
Volunteer tester
Avatar

Send message
Joined: 4 Apr 01
Posts: 201
Credit: 47,158,217
RAC: 0
United Kingdom
Message 1003510 - Posted: 12 Jun 2010, 7:36:47 UTC

I have seen a few people stating that they are running 2 tasks on their gtx470&480's - why? When I look at gpu usage, I typically see 69% usage peaking in the 90's for short bursts. Is there any benefit to doing this, would it actually work out any quicker as Im thinking that it might crunch 2 units but would do so at half the speed? Am I right?
ID: 1003510 · Report as offensive
Profile zoom3+1=4
Volunteer tester
Avatar

Send message
Joined: 30 Nov 03
Posts: 66286
Credit: 55,293,173
RAC: 49
United States
Message 1003518 - Posted: 12 Jun 2010, 7:49:56 UTC - in response to Message 1003510.  
Last modified: 12 Jun 2010, 7:55:51 UTC

I have seen a few people stating that they are running 2 tasks on their gtx470&480's - why? When I look at gpu usage, I typically see 69% usage peaking in the 90's for short bursts. Is there any benefit to doing this, would it actually work out any quicker as Im thinking that it might crunch 2 units but would do so at half the speed? Am I right?

Based on experience with the GTX295 and sometimes running the thing with 4 WU's, So I'd say No and agree 1 WU per gpu is preferred.
Savoir-Faire is everywhere!
The T1 Trust, T1 Class 4-4-4-4 #5550, America's First HST

ID: 1003518 · Report as offensive
Previous · 1 · 2 · 3 · 4 · 5 · 6 . . . 15 · Next

Message boards : Number crunching : Running SETI@home on an nVidia Fermi GPU


 
©2024 University of California
 
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.