karmic: stuck

[BeemerBiker]

Cannot seem to get crunch3r's seti cuda working.  I got it working on Dotsch_UX (8.1) but am doing something wrong on 9.1

ldd shows the cufft and cuda lib's and AFAICT i set them up correctly as shown below




[edit] I grabbed a screen print of the ldd output from Crunch3r's working seti cuda from my lDotsch_UX box.  I dont see much difference

[sunu]

I don't see a corresponding ldd pic from 9.10.

[BeemerBiker]

Scroll the window to the left to see the ldd results.

However, I am still looking at the problem.  I installed the linux cuda toolkit, the 2.3 version.  While that diid not (yet) get seti cuda working, it did get me a nice=0 for collatz.  Previously, the linux cuda 2.02 versoin of collatz was running at nice=19 which was unable to feed itself.  The 2.3 library replaced the 2.2 that came with that 2.02 download.   I assume that replaceing the 2.2 with the 2.3 made the "nice" difference.   FWIW, you can look here to see the collatz problem I was having.   I got  that seti cuda linux version working on 8.1 a week ago and it should work on 9.1.  I am still getting that url download error, but the ldd now shows I am using the 2.3 libraries  Anyway, tomorrow is another day and it will eventually work.

[sunu]

First of all get the newer setiathome-6.08.CUDA_2.2_x86_64-pc-linux-gnu from http://calbe.dw70.de/mb/viewtopic.php?f=9&t=116 .

What is exactly your problem? That boinc wants to download those cuda libraries? Just select them and abort them.

[Pepi]

You can download  cuda toolkit 3.0 from nvidia page and linux cuda 3.0 application  from here http://calbe.dw70.de/mb/viewtopic.php?f=9&t=120
( it is x64)


Also I have same problem and look at this( this help me) now I crunch on cuda 3.0 with latest Dotsch_UX
http://lunatics.kwsn.net/linux/seti-mb-cuda-for-linux.msg25956.html#msg25956

[sunu]

Pepi, do you have a fermi card? If you don't, go back to cuda 2.2 app and cuda 2.3 libs. It is faster. Look at http://lunatics.kwsn.net/linux/setihomecuda-3-0-linux-apps.0.html .

[Pepi]

Sunu, I was reading your article in 4 AM ( my local time) and go back to Cuda 2.2 APP and cuda libs 2.3.
But it looks like that combination is about 5 seconds slower then  cuda 3.0 and cudalibs 3.0 with my linux and my card  GTX 260. And since all results I now have is short ones I will wait unit one or more long appear, so I can compare old "long" and new "long" results.

[sunu]

Sunu, I was reading your article in 4 AM ( my local time) and go back to Cuda 2.2 APP and cuda libs 2.3.
But it looks like that combination is about 5 seconds slower then  cuda 3.0 and cudalibs 3.0 with my linux and my card  GTX 260. And since all results I now have is short ones I will wait unit one or more long appear, so I can compare old "long" and new "long" results.

You can't compare directly 2-3 different short or long ones as you say, run through boinc, and decide which is faster. Every workunit is different. You have to run the SAME workunits with both combinations to see which is faster, like I did in the other post. If you want to set up such a test and need help post here.

[Pepi]

Of course I need help, but my knowledge of linux is very very low. So  if you still wont to help me: please say: how to do tests?
I am interested in only two combination: cuda 2.2 app + cuda 2.3 dll and cuda 3.0 app + cuda 3.0 dll

[sunu]

Ok.

1. Get the file I attach below and extract it somewhere.

2. In the base folder KWSN... put the two cuda applications.

3. Go to your boinc cache and randomly pick some workunits and copy them to the testWUs directory.

4. Open a console in the testWUs directory and do a ls -1 to get a list of the workunits.

5. Edit the file benchmark and where it says FILES="" put the list from 4. between the two ", something like:
FILES="07mr07ae.3628.26592.15.10.150
29ja07ad.1241.892.5.10.236
29mr07af.21746.25021.16.10.234"
Where it says REF_APP="" put the filename of the 2.2 app between the two " and where it says APPS="" put the filename of the 3.0 app.

6. Set up the cuda 2.3 and cuda 3.0 libs side by side. The cudart.so and cufft.so symbolic links are the same between the two sets and could cause some confusion. I think you can omit them, at least for this test.

7. Close boinc and any other program you have running in the background. It is better to leave your PC alone while running this test to not skew the results in any way.

8. Run the benchmark file with
sh benchmark

9. In the end, in the testData folder  there will be a file named time.txt where you'll find the running times.

[attachment deleted by admin]

[Pepi]

THANKS!



My card is GTX 260
I am really appreciated your will and time you spend to write those instructions for me.
Since I have already have setup cuda 3.0 and cuda libs 3.0 I made benchmark for that. And since  I am not sure that I can setup both cuda libs I made two separate tests.

This is time for cuda 3.0 libs with cuda 3.0 app from crunch with Nvidia 195.36.15 drivers on DOTSCH_UX linux

01dc06ag.9607.45448.6.10.120./seti -verb  209.73 sec 53.13 sec 7.10 sec 28%
11ja07ai.1172.9070.7.10.107./seti -verb      718.98 sec 91.89 sec 19.22 sec 15%
13fe07ae.16925.4980.7.10.200./seti -verb  554.27 sec 78.45 sec 14.77 sec 16%
14fe07ae.16951.4162.6.10.233./seti -verb  808.03 sec 98.71 sec 20.74 sec 14%
15fe07ae.377.1708.3.10.66./seti -verb       650.18 sec 85.62 sec 16.91 sec 15%
27ja07ae.4631.28807.5.10.81./seti -verb    206.13 sec 52.15 sec 7.46 sec 28%
30dc06ah.479.17659.12.10.208./seti -verb 220.25 sec 53.33 sec 7.67 sec 27%
--------------------------------------------------------------------------------------------------------------   TOTAL
                                                       3367,57          513,28      93,87

This is time for cuda 2.3 libs with cuda 2.2 app from crunch with Nvidia 195.36.15 drivers on  DOTSCH_UX linux


01dc06ag.9607.45448.6.10.120./seti22 -verb  199.92 sec 54.08 sec 7.80 sec 30%
11ja07ai.1172.9070.7.10.107./seti22 -verb      709.99 sec 97.35 sec 19.13 sec 16%
13fe07ae.16925.4980.7.10.200./seti22 -verb   542.54 sec 82.44 sec 15.70 sec 18%
14fe07ae.16951.4162.6.10.233./seti22 -verb   797.42 sec 102.65 sec 21.41 sec 15%
15fe07ae.377.1708.3.10.66./seti22 -verb         637.19 sec 90.96 sec 17.18 sec 16%
27ja07ae.4631.28807.5.10.81./seti22 -verb     205.41 sec 53.45 sec 8.00 sec 29%
30dc06ah.479.17659.12.10.208./seti22 -verb  222.38 sec 55.30 sec 8.03 sec 28%
-------------------------------------------------------------------------------------------------------------
TOTAL                                                       3314,85         536,23       97,25

Difference is less then 2%

This is time for cuda 2.3 libs with cuda 2.2 app from crunch with Nvidia 195.36.15 drivers on  UBUNTU 10.04 BETA2

01dc06ag.9607.45448.6.10.120./seti -verb    206.14 sec 53.59 sec 10.86 sec 31%
11ja07ai.1172.9070.7.10.107./seti -verb        722.16 sec 97.39 sec 29.06 sec 17%
13fe07ae.16925.4980.7.10.200./seti -verb    560.39 sec 85.33 sec 23.43 sec 19%
14fe07ae.16951.4162.6.10.233./seti -verb    814.39 sec 103.99 sec 32.85 sec 16%
15fe07ae.377.1708.3.10.66./seti -verb   553.05 sec 81.65 sec 23.01 sec 18%
27ja07ae.4631.28807.5.10.81./seti -verb     210.92 sec 54.53 sec 11.11 sec 31%
30dc06ah.479.17659.12.10.208./seti -verb  227.31 sec 55.91 sec 11.91 sec 29%
----------------------------------------------------------------------------------------------------------------------------------------------------------------------
                                                             3294.36       532,4      142.23

By crunch time latest combination is fastest, but if you look one one result is really faster  (bold) all other are slower

[sunu]

Nice one Pepi. -1.59% to be exact 

So the combination cuda 3.0 app + cuda 3.0 libs give better results in a faster GPU like your GTX260 than a lower card like my 9600GT but still cuda 2.2 app + cuda 2.3 libs is faster.

Can you also post your results in Crunch3r's forum in the thread http://calbe.dw70.de/mb/viewtopic.php?f=9&t=120 ?

[Pepi]

Yes I will post it

But I cannot register, something is broken on forum

[sunu]

What problem do you have?

[Pepi]

Fixed, and posted!