V10/11 of modified SETI MB CUDA + opt AP package for full multi-GPU+CPU use

[Jason G]

...
What happened to the auto installer.
...

Over in the beta development thread,
http://lunatics.kwsn.net/public-release-beta-forum/windows-32-bit-only-installer-nt-xp-vista-win7.msg17174.html#msg17174

we need more information from you to be of help.

Jason

[Eithanius]

How should I edit the app_info so that I can crunch optimised AP via CPU and stock AP via CUDA...? No MB units involved. I'm not good at XML editing.

Thank you.

[Raistmer]

How should I edit the app_info so that I can crunch optimised AP via CPU and stock AP via CUDA...? No MB units involved. I'm not good at XML editing.

Thank you.

There is no CUDA AP available.

[Eithanius]

There is no CUDA AP available.

Then why is ap_graphics_5.03_windows_intelx86.exe being downloaded on my other instance of BOINC...?

I was wondering if this said AP client can be integrated into the existing V10 package. If so, how would I edit the app_info.xml to identify and utilise the client...?

Thanks

[Josef W. Segur]

There is no CUDA AP available.

Then why is ap_graphics_5.03_windows_intelx86.exe being downloaded on my other instance of BOINC...?

I was wondering if this said AP client can be integrated into the existing V10 package. If so, how would I edit the app_info.xml to identify and utilise the client...?

Thanks

ap_graphics_5.03_windows_intelx86.exe does not crunch WUs, it's the executable launched by the screensaver or BOINC Manager to display graphics. It is separate from the crunching app due to the way BOINC 6.x deals with Vista (and presumably Win 7) requirements.
                                                                             Joe

[Eithanius]

ap_graphics_5.03_windows_intelx86.exe does not crunch WUs, it's the executable launched by the screensaver or BOINC Manager to display graphics. It is separate from the crunching app due to the way BOINC 6.x deals with Vista (and presumably Win 7) requirements.
                                                                             Joe

oops...! my bad... thanks for knocking some sense out of me... 

[gerga]

Hi @ll

How's the best way to run MB and AP CPU opt app and MB CUDA opt app?

I mean, where do i download the opti applications and the corresponding app_info file? Should I do my own app_info file? Do u recommend version 6.6.20?

Thanks in advance.

[Claggy]

Hi @ll

How's the best way to run MB and AP CPU opt app and MB CUDA opt app?

I mean, where do i download the opti applications and the corresponding app_info file? Should I do my own app_info file? Do u recommend version 6.6.20?

Thanks in advance.

See this thread:

app_info for AP500, AP503, MB603 and MB608

Raistmer's app's can be found spread out in this thread,
the easiest way to get all the other optimised apps is to get the right combined package from Arkayn's Site:

Multibeam and Astropulse Combined Packages

Claggy

[gerga]

Hi @ll

How's the best way to run MB and AP CPU opt app and MB CUDA opt app?

I mean, where do i download the opti applications and the corresponding app_info file? Should I do my own app_info file? Do u recommend version 6.6.20?

Thanks in advance.

See this thread:

app_info for AP500, AP503, MB603 and MB608

Raistmer's app's can be found spread out in this thread,
the easiest way to get all the other optimised apps is to get the right combined package from Arkayn's Site:

Multibeam and Astropulse Combined Packages

Claggy

Thx for the info Claggy.

[FalconFly]

Hmpf...

I've downloaded the Package : Raistmer's_opt_pack_V10a_x86_SSE3_AMD_MB_CPU_GPU_team.rar

For WinXP, AMD Phenom 9550 und a GTX285 should be the right combination.

However, since I relaunched BOINC 6.6.20, the only Message I receive is :
Message from server: No work sent
Message from server : No work is available for Astropulse

At this time, however, there are 24592 AP WorkUnits and 120065 MB WorkUnits ready to send.

Setting the cc_config.xml to 5 CPUs (QuadCore + 1GPU) only resulted in the CPU attempting to crunch 5 AP WorkUnits (were already in Cache) at once...
The above Server Messages were the same with ncpus set to 5.

My Preferences allow for all Applications and GPU is set to be used even when Computer is in use.

The app_info.xml is unmodified right from the Package.

Any Ideas ?

-- edit --
For my single GPU config, I downloaded the stock CUDA application and manually added it into my old app_info.xml.
BOINC immediately downloaded MB WorkUnits and appears to work just fine.

Do I understand correctly that the stock CUDA Application does not support Multi-GPU configs (plan to build 2xGTX295 with 4 GPUs in a system) ?

[Josef W. Segur]

...
Message from server: No work sent
Message from server : No work is available for Astropulse

At this time, however, there are 24592 AP WorkUnits and 120065 MB WorkUnits ready to send.
...

Astropulse is the old application, it is quite likely there is no work available for that. Astropulse v5 is the new one, those 24592 tasks should be all or almost all for that application. Those server messages are exact, they'll say "Astropulse v5" if that is what isn't available.

However, the "no work is available" message refers to the 100 slots in the shared memory between the Feeder and Scheduler. At any moment, there may well be no work of a particular kind there, and the Scheduler can't send what it doesn't know about.

If the server message only reports about "Astropulse", that may mean your project preferences have somehow been set to only accept that kind of work.
                                                                              Joe

[Raistmer]

CUDA 2.2 benchmarking results:

1) New CUFFT can't run with 182.50 driver
2) new (2.2) CUDA runtime+ current CUDA MB can operate with 182.50 driver
3) there is small speedup from using CUDA_RT of 2.2 and there is no speedup from rebuilding current CUDA MB sources with CUDA 2.2
4) there is significant speedup (it even offsets lack of OCing on my GPU) from using current drivers (185.85) + CUDA RT 2.2 + CUFFT 2.2 (again, all builds show that speedup, not only freshly rebuilt one).

Summary: for those who can go with new drivers it's worth to upgrade to CUDA 2.2 runtime w/o upgrading current CUDA MB app.
For all others it's worth to try to use new cudart.dll  + old cufft.dll.

(new CUDA 2.2 runtime attached (cudart.dll + cufft.dll )

Don't forget to update video driver to 185.85 version.

[attachment deleted by admin]

[KarVi]

Thanks for the heads up Raistmer.

I´m trying the newest drivers (185.85), with CUDA 2.2 supplied by you, on my 9500.

Looking forward to see the results.

[FalconFly]

Astropulse is the old application, it is quite likely there is no work available for that. Astropulse v5 is the new one, those 24592 tasks should be all or almost all for that application. Those server messages are exact, they'll say "Astropulse v5" if that is what isn't available.

However, the "no work is available" message refers to the 100 slots in the shared memory between the Feeder and Scheduler. At any moment, there may well be no work of a particular kind there, and the Scheduler can't send what it doesn't know about.

If the server message only reports about "Astropulse", that may mean your project preferences have somehow been set to only accept that kind of work.
                                                                              Joe

Alright, so I will give it another shot and add more patience

I've read that BOINC 6.6.20 does not work with the V10, so which BOINC version is suitable ?
(Note - I'm using the Windows 7 RC 64bit, not sure if that affects the whole thing)

[Richard Haselgrove]

CUDA 2.2 benchmarking results:

1) New CUFFT can't run with 182.50 driver
2) new (2.2) CUDA runtime+ current CUDA MB can operate with 182.50 driver
3) there is small speedup from using CUDA_RT of 2.2 and there is no speedup from rebuilding current CUDA MB sources with CUDA 2.2
4) there is significant speedup (it even offsets lack of OCing on my GPU) from using current drivers (185.85) + CUDA RT 2.2 + CUFFT 2.2 (again, all builds show that speedup, not only freshly rebuilt one).

Summary: for those who can go with new drivers it's worth to upgrade to CUDA 2.2 runtime w/o upgrading current CUDA MB app.
For all others it's worth to try to use new cudart.dll  + old cufft.dll.

(new CUDA 2.2 runtime attached (cudart.dll + cufft.dll )

Don't forget to update video driver to 185.85 version.

I've applied the full 'triple treatment' (driver, runtime and fft) to my rigs (2 x 9800GT and a 9800GTX+), which run the stock v6.08 from Berkeley. Good impression all round: display seems smoother, app runs faster, and I'm even seeing signs of lower CPU loading (though maybe that's just BOINC's way of telling me the boxes were due a reboot ). I'm still not going to risk throwing a VLAR into the mix, but apart from that we seem to have a winner.