Modified SETI MB CUDA + opt AP package for full GPU utilization

[sunu]

I don't know if this will helps you

http://boinc.berkeley.edu/wiki/Anonymous_platform

Thanks, that was what I was looking for. Unfortunately I think I've already used

<avg_ncpus>3.0000</avg_ncpus>
<max_ncpus>3.0000</max_ncpus>

for AP with not the desired effect. Maybe I'll try again.

[Raistmer]

use <ncpus>Ncores+1</ncpus> in cc_config.xml instead. IT works much better than fiddling with CPU shares.

[sunu]

Raistmer here's the thing: I want to force 3+1 in a quad (not 4+1).

When I use <ncpus>4</ncpus> in cc_config.xml, boinc runs 4+1 when it is in normal mode and 3+1 in high priority mode.
When I use <ncpus>3</ncpus>, boinc runs 3+1 normal and 2+1 high priority.

All these are in boinc linux. Maybe it is a bug/feature?

[Raistmer]

then try 2. If you don't care about idle cores it should work. 

[Jason G]

Or you can try this: change the amount of cpus for the GPU app in the app_info file from 0.04, to 1  (or even 2  if needed)  then boinc can't allocate those cores to something else.  You might also want to check the flops setting in there is higher than for the CPU app (You can add #cpus (i.e. 1) and flops to CPU app section if it isn't there), so GPU app gets allocated first and hogs the required number of cores (x1)  prior to CPU apps being allocated(x remaining 3).

Jason

[sunu]

from 0.04, to 1  (or even 2  if needed)
Jason

I've never tried 2. I was stopping to 1 

New ideas, thanks!

[Jason G]

You could even try 1.04, if 2 cuts back too much,  that may play nicer with the newer CPU juggling with EDF mode.

[sunu]

Thanks Jason.

Regarding your previous post, I didn't know the trick with flops. Can I put just an arbitrary number?

[Jason G]

Yep, I gather it's just relative, but I would look how much GFLops in BoincView CPU app acheives, put that value (in Flops) for the CPU, then multiply by 2-5 (as appropriate for the card speed against CPU) for the GPU version.

Jason

[sunu]

I can't use boincview (linux). I've found some flop numbers in client_state.xml and I'll play with them.

[Slawek]

if some one want here is cufft and cudart  dlls from last CUDA 2.1 final :

stop boinc unpack and replace in boinc data dirs the old  dlls

Any imp. computing  ?

[sunu]

Just a FYI. Whatever I tried, flops, avg_ncpus, maxncpus in app_info.xml in order to force 3+1 in a quad didn't have any effect. The only thing that seems to work is

<avg_ncpus>1.3400</avg_ncpus>
<max_ncpus>1.3400</max_ncpus>

in AP section of app_info.xml. But that possibly means that If there is no MB work, only 3 cores will be used. Whatever I did in MB section seemed irrelevant.

[Jason G]

...
in AP section of app_info.xml. Whatever I did in MB section seemed irrelevant.

Ahhh, of course, I think in 6.4.5 or higher they might have made it ignore the Cuda CPU component! LoL. Madness 

[ceciltseng]

hi all ,
AP5.0 SSE4.1  + CUDA MB possible to use now?   ( 64 bit version is wonderful )

[Raistmer]

hi all ,
AP5.0 SSE4.1  + CUDA MB possible to use now?   ( 64 bit version is wonderful )

There is no SSE4.1 AP at all, did you mean AK_v8 SSE4.1 ? I found not improvement in speed for SSE4.1 over SSSE3x version. In multicore system SSSE3 performs better actually (but difference pretty slow anyway). So I don't plan to do separate SSE4.1 version. Also, there is no AP x64 build for now.