ABP2 CUDA applications

Anonymous
Topic 13801

Quote:
Any news on ABP2 CUDA? or a 'possible' ATI client?


The initial ABP2 CUDA App will be essentially the same as the ABP1 one, i.e. it will just run the FFT on the GPU and the other calculations on the CPU. A custom kernel to bring these to the GPU, too is still being worked on.

If there is a similar library to cuFFT for ATI one could use the same approach, but with the same drawback (lots of computation still done on the CPU).

Currently it looks like we'll finish the CUDA kernel for ABP2, the port this to an OpenCL version that would support both GPUs (and then some computing devices).

BM

Bernd Machenschalk
Bernd Machenschalk
Joined: 15 Oct 04
Posts: 2,684
Credit: 25,950,161
RAC: 34,820

ABP2 CUDA applications

Quote:

Just happened to notice this:

einsteinbinary_ABP2_5.02_i686-apple-darwin__ABP2cuda23

Does that mean CUDA app is now available on the OS X platform...?


In principle CUDA has been available for Mac OS X for a while.

There were and probably are problems in conjuntion with the BOINC Core Client, that look like been resolved a few hours ago. Einstein@home CUDA Apps could work with the latest development client for Mac OS X installed without sandboxing or a custom Client built from BOINC SVN trunk HEAD.

In any case manual installation of the latest NVida CUDA driver is required, I think it's only available for SnowLeopard.

BM

BM

Bernd Machenschalk
Bernd Machenschalk
Joined: 15 Oct 04
Posts: 2,684
Credit: 25,950,161
RAC: 34,820

RE: Einstein@home CUDA Apps

Message 3727 in response to message 3726

Quote:

Einstein@home CUDA Apps could work with the latest development client for Mac OS X installed without sandboxing or a custom Client built from BOINC SVN trunk HEAD.

In any case manual installation of the latest NVida CUDA driver is required, I think it's only available for SnowLeopard.

Update: The 6.10.28 Client of Jan 12 from http://boinc.berkeley.edu/download_all.php seems to work even with sandboxing.

BM

BM

Bernd Machenschalk
Bernd Machenschalk
Joined: 15 Oct 04
Posts: 2,684
Credit: 25,950,161
RAC: 34,820

We'll reduce the limit from

We'll reduce the limit from 450 to 400 or 420. This requires replacing the scheduler. I'm afraid I can't do this today, first thing tomorrow.

BM

BM

Bernd Machenschalk
Bernd Machenschalk
Joined: 15 Oct 04
Posts: 2,684
Credit: 25,950,161
RAC: 34,820

RE: We'll reduce the limit

Message 3729 in response to message 3728

Quote:
We'll reduce the limit from 450 to 400 or 420. This requires replacing the scheduler. I'm afraid I can't do this today, first thing tomorrow.

Done.

I suggest to abort the affected tasks and update the project to get new GPU tasks.

BM

BM

Bernd Machenschalk
Bernd Machenschalk
Joined: 15 Oct 04
Posts: 2,684
Credit: 25,950,161
RAC: 34,820

RE: Any news on an updated

Quote:
Any news on an updated ABP2 CUDA version with better GPU utilization, or on the ATI version?

A better CUDA ABP2 App is being worked on, but we still have trouble with numerical differences between GPU and CPU calculations. Oliver will be on well deserved vacation for the next weeks, and I'll focus on preparing the next GW search, so there won't be much progress on CUDA in February.

I doubt that we'll ever make a native ATI App, we may at some point have OpenCL Apps that will run on ATis, too.

BM

BM

Bernd Machenschalk
Bernd Machenschalk
Joined: 15 Oct 04
Posts: 2,684
Credit: 25,950,161
RAC: 34,820

RE: Is the Spy-Hill

Quote:
Is the Spy-Hill workround still needed?


AFAIK No. The Client uses only a single Shared Memory segment to communicate with the Apps; the communication between App and Graphics should be implemented with memory mapped files.

BM

BM

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.