AK V8 + CUDA MB team work mod

[Vager]

Ok here's the reason why my GPU stops working.. I'm not sure why..

Here's what a normal completed task looks like when it's worked on by my GPU

setiathome_CUDA: Found 1 CUDA device(s):
   Device 1 : GeForce 8800 GTS
           totalGlobalMem = 335544320
           sharedMemPerBlock = 16384
           regsPerBlock = 8192
           warpSize = 32
           memPitch = 262144
           maxThreadsPerBlock = 512
           clockRate = 1350000
           totalConstMem = 65536
           major = 1
           minor = 0
           textureAlignment = 256
           deviceOverlap = 0
           multiProcessorCount = 12
setiathome_CUDA: No device specified, determined to use CUDA device 1: GeForce 8800 GTS
SETI@home using CUDA accelerated device GeForce 8800 GTS
CPU-GPU team mod by Raistmer
Priority normalize modification by Raistmer
Priority of worker thread rised successfully
Priority of process adjusted successfully
Total GPU memory 335544320    free GPU memory 230916096

Now here's what happens after SETI is CPU-GPU team is running for maybe 12 to 24 hrs.

setiathome_CUDA: Found 1 CUDA device(s):
   Device 1 : Device Emulation (CPU)
           totalGlobalMem = -1
           sharedMemPerBlock = 16384
           regsPerBlock = 8192
           warpSize = 1
           memPitch = 262144
           maxThreadsPerBlock = 512
           clockRate = 1350000
           totalConstMem = 65536
           major = 9999
           minor = 9999
           textureAlignment = 256
           deviceOverlap = 0
           multiProcessorCount = 16
setiathome_CUDA: device 1 is emulation device and should not be used, supports 9999.9999
setiathome_CUDA: No SETI@home capabale CUDA GPU found...
SETI@home NOT using CUDA, falling back on host CPU processing

The totalGlobalMem is showing -1 and the Major Minor values are showing 9999, 9999 . I also notice that the MB_6.08_mod_CPU_team_CUDA.exe Process is pegged at 50% instead of the usual low value everytime the CPU feeds data to the GPU. Then What I do to simply fix this is just restart BOINC, after that everything seems to run fine once again.

Please keep in mind that I just today upgraded to a Quad Core Q6600, so I'm not sure if this will still happen. I was running previously a Dual Core E6600.

Thanks

[Archangel999]

i use Asus Striker 2 Extreme but my memory is weak it is only 1333mhz corsair  if i had a 1800 wow 

Best Regards
D.Draganov

[Raistmer]

setiathome_CUDA: device 1 is emulation device and should not be used, supports 9999.9999
setiathome_CUDA: No SETI@home capabale CUDA GPU found...
SETI@home NOT using CUDA, falling back on host CPU processing

Program failed to find your GPU for some reason... try to chnage video-drivers or underclock GPU a little - maybe it will help...

[Raistmer]

I lost main HDD drive on desktop PC tomorrow (it fails on self-boot test, some hardware failure) so will be busy with data restoration for some time...

[k6xt]

I don't know if this is believable: My PC has run Raistmer's optimization for several days now. Some of the CUDA units are earning results. Comparing work units earning similar credit (i.e. say 50 points) a CPU will take 3500 seconds while the GPU takes 130 seconds.

This would mean my GPU 8500GT, one of the slowest  and oldest CUDA capable on Nvidias list, without any overclocking, processes about 7 times faster than all four overclocked Q6600 cores combined!

But, there are several (4 to 8 or more) CPU work unit results  between each GPU result in the PC task list. This fact would argue at best the GPU is only equal to one CPU. Or that the GPU is frequently idle. I don't quite know how to find out.

Maybe the times are not accurate.

[Raistmer]

It was disscussed many times already. Look on elapsed times, not CPU ones....

[k6xt]

It was disscussed many times already. Look on elapsed times, not CPU ones....

Where do I find 'elapsed time'? I do not see it in the task list area. I went back through stdoutdae.txt, cannot find any task taking under one hour. Am I in the wrong file?

[Raistmer]

in stderr of finished result.
On web site go to result info

[Stelminator]

just wanted to mention, Windows 7 is less than useful for this trick.

I ran this on Windows 7 64bit.  Definitely stupid.  The drivers are old, 179.23 (GPUZ ftw).  From other threads, I assume 180 or something higher is required for this to function properly.  (I figured that out after I realized something was wrong).

That said, I believe the teamwork algorithm was working superbly.  Other than the fact that every CUDA WU errored out almost immediately, there were exactly the right number of tasks for my i7 system; 9 total, with 8 maxing the CPU, and one more only stealing part of 1 cpu long enough to do setup or teardown.  No idea how much it would use while functioning, because they all died quickly.  Also, 2 of the cpu tasks were AP's.

Here's hoping BOINC will fix the scheduling sooner rather than later, or Nvidia will release new drivers for me.  Unfortunately, there's no way I'm going back to Vista.  Win7 is "Vista, done right", imo.

if/when there are drivers that will let this work, is there any chance of a SSE4.1 build?  On that note, I haven't been able to find what the speed difference for SSE4.1 vs SSSE3, SSE3, and so forth are, nor anything regarding if SSE4.2 provides anything useful to the project.  If/when I'm able, I'll certainly experiment to see if 8x SSSE3 on i7 + 1x 9400 GT is faster than just 8x SSE4.1.

Great job and thank you!

[NudgeyNR]

My system....
Q6600@2.66
4GB DDR2 1066
GTX260
8600GTS
Boinc V6.6.0

Any ideas why boinc is reporting my system has a GTX260 & 8600GTS, but my account data on berkeley is saying 2 x 8600GTS's ?
I am using the V8 + CUDA apps with great success, how do i know which GPU is being used ?

Also i have tested Boinc 6.6.4 without the AK apps and it runs 2 CUDA wu's plus 4 CPU wu's noproblems. The CUDA wu's are extremely fast, just wish i could use the opt apps along side with the same success. I tried running the AK V8 + CUDA apps on 6.6.4 but Boinc crapped out and would not display any stats even though the wu's were still being processed.

[k6xt]

in stderr of finished result.
On web site go to result info

Would that be this statement? If so I don't understand the stated time.

"Wall-clock time elapsed since last restart: 6902.6 seconds
called boinc_finish"

Also the above statement doesn't seem to exist in all completed WU that have been granted credit. Is this statement a part of your work and applies only to GPU-crunched WU?
Sorry for the bandwidth....just trying to understand what is happening.

[Yellow_Horror]

How you like an idea to combine "VLAR stays on CPU" with "Advance to free GPU" techniques?

For example, let 3 WU's are crunched at 2core+1GPU system. The GPU task is finished up, take the new one - oops it's VLAR - stays on the CPU. For some time 3 WU's are crunched on CPU, then one of the previously running non-VLAR tasks will detect free GPU and switch to it.

Yes, it can end up in 3 VLARs crunching on CPU while GPU stays idle. But with more cores the situation is less probable. The main advantage of this scheme is that VLAR is never end up on the GPU giving the lazy uzer interface.

I am interested to test such version if you decide to create one (my CPU is limited to SSE3).

[Raistmer]

in stderr of finished result.
On web site go to result info

Would that be this statement? If so I don't understand the stated time.

"Wall-clock time elapsed since last restart: 6902.6 seconds
called boinc_finish"

Also the above statement doesn't seem to exist in all completed WU that have been granted credit. Is this statement a part of your work and applies only to GPU-crunched WU?
Sorry for the bandwidth....just trying to understand what is happening.

It's elapsed time value. From lats task restart or from beginning of task if there  were no restarts.
This output now added o all  new "team" builds.

[Richard Haselgrove]

Heads-up for people testing out the latest Alpha BOINC clients, and fiddling with cc_config.xml settings.

The setting of <ncpus>0</ncpus> is NO LONGER the way to set it back to whatever the computer has (and I have this as a place holder on several machines).  It has to be <ncpus>-1</ncpus> to work.  Perhaps a warning about this should be put in the config documentation page is it is a rather nasty and hard to figure out problem.

That's right; I recently changed things so that <ncpus>0</ncpus> means use zero CPUs (this was in response to a request from someone who wanted to compute only with GPU).
To get the actual #CPUs, use -1 or omit the element.

[Archangel999]

Raismer there is some kind of bug i use the ak v8 + cuda mb mod and cc_config file to crunch 5 at a time work unit but at some time after working it is stoped crunch the 5-th unit it the boinc it says it is runnig but no progreess and the cpu is not at 100% after that  it is 75% again and after some time it goes only 50% and cunching only 3 work unit the other 2 says again in boinc that they are runnig but no progress when i go to the task manager they are in the memory but maybe they are stuck ?

or mayby have to start the cuda log but it is for cuda only here they are a cpu applications
and cuda log work perfect but the app that freezze is AK_v8b_win_SSSEx_GPU_CPU_team.exe*32 but why is 32 bit will you do 64bit for windows like mine xp x64 ??
and with some time the windows has prompt me a error message from this application when it stuck
Best Regards
D.Draganov



[attachment deleted by admin]