SETI MB CUDA for Linux

[Hefto99]

hmm... I have tried  different environments (gnome, kde 3.5, xfce..) , but no improvement...
It seems I have to wait for the better GK.. 

Thanks to everybody for support!

[koschi]

I'll check the download sector and come back to the thread when I have some benchmark results ;-)

[Hefto99]

OK, I have finally got it running under  the XFCE display manager 
14.21 credit WU has been finished in under 15 minutes, although it is still falling fallback:
  CUDA runtime ERROR in device memory allocation (Step 1 of 3). Falling back to HOST CPU processing...

[Raistmer]

OK, I have finally got it running under  the XFCE display manager 
14.21 credit WU has been finished in under 15 minutes, although it is still falling fallback:
  CUDA runtime ERROR in device memory allocation (Step 1 of 3). Falling back to HOST CPU processing...

It means you actually use 6.03 stock app CPU code for processing taks. If you can't avoid fallback to CPU it's much better to use one of AK v8 opt apps for CPU processing than 6.08 in CPU mode.

[Hefto99]

Other tasks look much better, no more runtime ERRORS 

[Raistmer]

Other tasks look much better, no more runtime ERRORS 

And CPU usage with that tasks ?

[Hefto99]

CPU usage is still high - 95-100 %
GPU temp is 15 C above normal.

[Raistmer]

with increased GPU temp it seems GPU works now.
Interesting what runtimes will be (wall clock ones)

[Claggy]

I had a look at Hefto99's http://setiathome.berkeley.edu/results.php?hostid=4774614tasks earlier, they are all using the GPU now,
with only a couple that did CPU fallback, but either finished shortly after, or restarted on the GPU, times are now in the thousands range,
instead of tens of thousands, about the speed of my T8100 C2D using AK_V8.

Claggy

[Raistmer]

CPU speed?
Well, CPU time for Win CUDA MB ~3-4 minutes.
And I more interesting in wall clock time....
Could someone with Linux CUDA run some of test WUs like PGxxx.wu or testwu_x.wu ?

[Hefto99]

My CPU is running on 2GHz (Athlon X2 3800+), GK is 8600 GT on default clocks, openSUSE 11.1 64-bit, here are some results:

http://setiathome.berkeley.edu/result.php?resultid=1142061689
CPU time:  883.97 sec.
Wall clock: 1087 sec.
Claimed credit: 14.20

http://setiathome.berkeley.edu/result.php?resultid=1141789428
CPU time: 2646 sec.
Wall clock: 2928 sec.
Claimed credit: 57.28

http://setiathome.berkeley.edu/result.php?resultid=1141486869
CPU time: 4652 sec.
Wall clock: 4756 sec.
Claimed credit: 81.32

CPU utilization is almost 100% for SETI
I will try to run some test units during next week.

[koschi]

downloaded http://lunatics.kwsn.net/index.php?module=Downloads;sa=dlview;id=18, results are like this:

root@cruncher:~# for i in 1 2 3 4 5 6 7 ; do cp testWU-$i work_unit.sah; rm stderr.txt; echo "starting testWU-$i"; time ./setiathome-CUDA-6.08.x86_64-pc-linux-gnu ; cat stderr.txt | grep "angle range"; echo;done
starting testWU-1

real    1m5.579s
user    1m3.372s
sys     0m0.216s
WU true angle range is :  0.604884

starting testWU-2

real    1m16.346s
user    1m14.065s
sys     0m0.276s
WU true angle range is :  0.443732

starting testWU-3

real    1m20.750s
user    1m18.493s
sys     0m0.268s
WU true angle range is :  0.425877

starting testWU-4

real    0m44.151s
user    0m41.959s
sys     0m0.184s
WU true angle range is :  1.279649

starting testWU-5

real    1m16.539s
user    1m14.309s
sys     0m0.232s
WU true angle range is :  0.439957

starting testWU-6

real    3m40.854s
user    3m38.670s
sys     0m0.200s
WU true angle range is :  0.033858

starting testWU-7

real    0m49.831s
user    0m47.635s
sys     0m0.192s
WU true angle range is :  0.775000

App was running with priority 0 and consuming 100% of one core, no other BOINC processes running. The system is a dedicated cruncher, C2Q6600 with GF9800GT.
If you need any other output from stderr.txt or result.sah, let me know

[Raistmer]

real, user and sys - are they CPU times ? try to use standardized KWSN testbench, it should show elapsed times too.

[koschi]

real is wall clock time, user is CPU time and sys is like kernel time or time spent interacting with the kernel, somehow like this...

I checked the KWSN testbench, it could need some small documentation, but I think I figured it out, its running now.
default-512 vs. AK_V8_linux64_ssse3 vs. setiathome-CUDA-6.08.x86_64-pc-linux-gnu

Will update this post when results are available...

edit:

I don't know what kind of work units these are, but they don't reflect what I've seen with current MB units.
The SSSE3 app is faster some times then my GPU in this bench. In real world performance on the project it is at least twice faster than the Q6600@3,2, except VLARs, of course... run times are attached in the text file...

my_science_app1 = AK_V8_linux64_ssse3
my_science_app2 = setiathome-CUDA-6.08.x86_64-pc-linux-gnu



[attachment deleted by admin]

[Raistmer]

Thanks.
Will do test with these WUs on my host under Windows and post for comparison.