SETI MB CUDA for Linux

[Raistmer]

The problem with Linux it cosumes 100% of CPU clock, not 2,5%....

[Jason G]

Sure, that's an implementation issue they'll need to work out, and maybe something about their cuda drivers & SDK  etc on that platform  (We know how well beta 2.1 SDK is refined on Windows after all.)

If fixable, I can reasonably expect that their apps would still report 1.5->2x the CPU time as the windows ones, since the Windows one has that hidden component.

 [ Edit: Without knowing about how Linux schedules its processes, It's probably implemented with spin-wait polling  loops instead of interrupts & DPCs, so reports 100% CPU, even though only some fraction of that is used for useful work.  in a spin-wait, it shouldn't be blocking other threads or resources anyway, so it shouldn't really prevent any other app running etc... even though it says 100%.

 It would be interesting to know the behaviour on Linux when a CPU wants to use that core, if the Cuda app yields most of the percentage or not. ]

[Raistmer]

About hidden component - interesting to measure it more accurately.
I understood that I could miss it in my tests (I measured only single core, other was used by BOINC) Will try to run 4 CPU apps + 1 GPU app all in standalone under appTimes control. This way all elapsed times will be accounted - no hidden component at all.

[Jason G]

Yeah difficult to nail down, as it won't show in wall or cpu time for that app. It will be small, but large in proportion to reported CPU time & consume about 50% on top of whatever reported cpu time is.  Only way I've been able to see this figure is in process explorer. 

next question is does it matter? Well no, because it is really hidden and doesn't add to wall time, but it does consume some small portion of overall machine resources available to other apps.

[Raistmer]

next question is does it matter? Well no, because it is really hidden and doesn't add to wall time, but it does consume some small portion of overall machine resources available to other apps.

And that portion should be known to do decisions about best host configs and performance tuning.
Overall host performance - that's really matter, not performance of just single app .

[Lysia]

I'm also seeing this

Cuda error 'GaussFit_kernel' in file './cudaAcc_gaussfit.cu' in line 506 : invalid configuration argument.

Beginning of stderr.txt is

Code: [Select]

SETI@home MB CUDA 608 Linux 32bit SM 1.0 - r06 by Crunch3r :p

setiathome_CUDA: Found 1 CUDA device(s):
   Device 1 : GeForce 8800 GTS 512
           totalGlobalMem = 536543232
           sharedMemPerBlock = 16384
           regsPerBlock = 8192
           warpSize = 32
           memPitch = 262144
           maxThreadsPerBlock = 512
           clockRate = 1620000
           totalConstMem = 65536
           major = 1
           minor = 1
           textureAlignment = 256
           deviceOverlap = 1
           multiProcessorCount = 16
setiathome_CUDA: CUDA Device 1 specified, checking...
   Device 1: GeForce 8800 GTS 512 is okay
SETI@home using CUDA accelerated device GeForce 8800 GTS 512
setiathome_enhanced 6.01 Revision: 402 g++ (GCC) 4.1.2 20070925 (Red Hat 4.1.2-33)
libboinc: BOINC 6.5.0

Work Unit Info:
...............
WU true angle range is :  0.447697
Optimal function choices:
-----------------------------------------------------
name               
-----------------------------------------------------
              v_BaseLineSmooth (no other)
   v_vGetPowerSpectrumUnrolled 626.74337 0.00000
             sse1_ChirpData_ak 49414.78434 0.00098
                 v_vTranspose4 22247.17831 0.00000
                BH SSE folding 9764.43373 0.00000
Cuda error 'GaussFit_kernel' in file './cudaAcc_gaussfit.cu' in line 506 : invalid configuration argument.
I'm running BOINC 6.4.5, Ubuntu 8.04 (2.6.24-23) and have a 8800 GTS 512 (G92). http://setiathome.berkeley.edu/show_host_detail.php?hostid=4840816
Nvidia driver is 180.29, I have tried with CUDA Toolkit from Nvidia (the versions of the libs are 2.1) and the CUDA libs delivered with the S@H-client, with the same results.

The only thing I seem to have in common with all the other reports in this thread seems to be that I run a 32 bit Linux.

I have calculated some work units, most of them still need validation, two are in state "Completed, validation inconclusive", so I guess the results may be garbage.

And some work units just gave a "Compute Error" (http://setiathome.berkeley.edu/result.php?resultid=1185431682), is this normal?

By the way:
Is there any good reason, why libcufft.so.2 and libcudart.so.2 are mentioned in app_info.xml? That should be normal libraries, and I don't think boinc should need to know about them. I did not unpack them at the beginning (had installed CUDA Toolkit), and boinc tried to download these files (and failed) now I have them in projects/setiathome.berkeley.edu but I don't think they are used from there.

[sunu]

Lysius, what kind of cuda client are you running? Those tasks marked "Error while computing" seem to have been killed with the VLAR autokill. As far as I know there hasn't been released a linux cuda client with a VLAR autokill function.

[Lysia]

Lysius, what kind of cuda client are you running? Those tasks marked "Error while computing" seem to have been killed with the VLAR autokill. As far as I know there hasn't been released a linux cuda client with a VLAR autokill function.

I'm using the Download from the first post in this thread (i686), redownloaded and verrified that this is really what I have installed. Is there anything newer?
And where can I find the source?

Has anyone successfully used the 32-bit version or does everybody use 64-bit?

[sunu]

Yes Lysius, there is something wrong with the 32bit app. I've found and downloaded one of the WUs that you've done ( http://setiathome.berkeley.edu/workunit.php?wuid=424266619 ). The 32bit app gives me the same errors, while the 64bit app is good. Worse still, the two results are weakly similar.

Note to Crunch3r: I've posted the two results in the development thread.

[Lysia]

Yes Lysius, there is something wrong with the 32bit app. I've found and downloaded one of the WUs that you've done ( http://setiathome.berkeley.edu/workunit.php?wuid=424266619 ). The 32bit app gives me the same errors, while the 64bit app is good. Worse still, the two results are weakly similar.

Any news on this?

And is there a guide anywhere on how to test an app without using BOINC and reporting possible garbage?

[sunu]

Any news on this?

Unfortunately no. All 32bit builds had the same error. If you can, run the 64bit app, else it's better not run CUDA for the time being.

And is there a guide anywhere on how to test an app without using BOINC and reporting possible garbage?

Just put  the app, a workunit named work_unit.sah and the file from the compressed archive I attached below in a directory and run it.

[attachment deleted by admin]

[Lysia]

Any news on this?

Unfortunately no. All 32bit builds had the same error. If you can, run the 64bit app, else it's better not run CUDA for the time being.

Then I think somebody should add a warning to the original post or remove the 32bit app entirely.

Unfortunately I only have 32bit linux installed. The last time I tried 64bit I had a sound problem in some 32bit games, so I just reverted to 32bit. Perhaps I should give it a try again, but I don't think this will happen anytime soon.

[Moustacha]

Excellent work. Gone from ~14,000 seconds to ~1,300 seconds just by using the GPU 

[kpolberg]

Is the source code for this available? Just wondering if it would be possible to compile a 32bit my self just to test. I would really like to put a 9500GT to work(my other htpc is based on xbmc with vdpau mod, would just like to try if crunching at htpcing is possible at the same time).

[Raistmer]

Is the source code for this available? Just wondering if it would be possible to compile a 32bit my self just to test. I would really like to put a 9500GT to work(my other htpc is based on xbmc with vdpau mod, would just like to try if crunching at htpcing is possible at the same time).

You could start with sources from Berkeley's SVN.
https://setisvn.ssl.berkeley.edu/svn/branches/seti_cuda
It contains stock CODA.