Parallella, Raspberry Pi, FPGA & All That Stuff

Bikeman (Heinz-Bernd Eggenstein)
Bikeman (Heinz-...
Moderator
Joined: 28 Aug 06
Posts: 3522
Credit: 728145867
RAC: 1194736

Thanks for the data point.

Thanks for the data point. And indeed the CHIP is a bit fatster. Both can run at 1GHy CLPU clock but the CHIP has a more modern CPU which supports the SIMD Neon instructions. Still I want to get a Pi Zero as well.

Cheers
HB

Tom Rinehart
Tom Rinehart
Joined: 17 Jun 09
Posts: 9
Credit: 6591748
RAC: 0

RE: There has been some

Quote:
There has been some work on GPU FFT reported here but nothing production-ready yet.

It looks like the software for the GPU FFT got updated last year:

http://www.aholme.co.uk/GPU_FFT/Main.htm

It looks like the accuracy was improved.

KF7IJZ
KF7IJZ
Joined: 27 Feb 15
Posts: 110
Credit: 6108311
RAC: 0

RE: It looks like the

Quote:


It looks like the software for the GPU FFT got updated last year:

http://www.aholme.co.uk/GPU_FFT/Main.htm

It looks like the accuracy was improved.

Please excuse my ignorance, as I'm new to doing anything other than running BOINC. Does this mean that BOINC is already using the GPU or capable of using the GPU? Is there something I need to do to enable this or has this yet to be implemented by the Einstein@Home folks?

My YouTube Channel: https://www.youtube.com/user/KF7IJZ
Follow me on Twitter: https://twitter.com/KF7IJZ

BackGroundMAN
BackGroundMAN
Joined: 25 Feb 05
Posts: 58
Credit: 246736656
RAC: 0

Hi, About a year ago I was

Hi,
About a year ago I was working on using the GPU_FFT library on the RPi B+ to accelerate the EaH client.
At first there are issues with the GPU_FFT accuracy on the 2M-point C2C FFT.
I communicate with Andrew Holme (GPU_FFT developer) and he sent me a version of the code with high accuracy for the 2M-point FFT.
Furthermore, Andrew add to the GPU_FFT a 4M-point C2C FFT kernel with high accuracy and I was able to minimize the calculations in the EaH client.
I think that the latest release of the GPU_FFT library includes these changes.

I managed to validate several WUs with my code.
The speedup on a RPi B+ (@1GHz,RAM@600MHz,GPU_MEM@216MB) with the use of GPU_FFT for the EaH is about 20-25% but there are some problems:
1. The EaH must be run with root privileges (unsecured). There is a library which probably allow you to run GPU_FFT as normal user but I was never test it.
2. On stopping the EaH client there are problems with the cleanup of the GPU and a reboot of the RPi is required.
3. After some days of running the EaH hangs without any obvious reason and a reboot is required.

After a while the RPi2 hit the market and the new processor was much faster than the GPU_FFT in FFT calculations (due to NEON).
Furthermore, due to quad-core processor if you need the GPU_FFT calculation you must run only one EaH client (low GPU memory for 4 clients, GPU sharing problems).
I didn't continue to resolve these problems.

The RPi Zero has the same SoC as RPi B+ so it is possible to run the EaH with the GPU_FFT with about 20-25% speedup, but with the above problems.

Tom Rinehart
Tom Rinehart
Joined: 17 Jun 09
Posts: 9
Credit: 6591748
RAC: 0

It looks like this project is

It looks like this project is working on some of the implementation problems like the root permission issue:

https://github.com/gpu-fftw/gpu_fftw

Chris S
Chris S
Joined: 27 Aug 05
Posts: 2469
Credit: 19550265
RAC: 0

RE: CPU/GPU core freqency

Quote:
CPU/GPU core freqency has it limits as does (22nm) process shrinking to moleculair level.

14Nm

10Nm

Waiting for Godot & salvation :-)

Why do doctors have to practice?
You'd think they'd have got it right by now

KF7IJZ
KF7IJZ
Joined: 27 Feb 15
Posts: 110
Credit: 6108311
RAC: 0

A question about wisdom

A question about wisdom files. For three of my PI 2 nodes, I am running on Wheezy and had custom compiled a BOINC 7.2.47 version by following the instructions in claggy's thread on the subject. I used a copy of the fftw wisdom file that was posted in this thread tied to fftw-3.3.2 and this is the result: https://einsteinathome.org/host/12039560/tasks. <50,000 second results.

Tonight, I added a new PI 2 to the cluster but it's based on the recently released Jessie Lite, which has BOINC 7.4.23. I had forgotten about the wisdom file, so I copied the one for 3.3.2 to /etc/fftw/wisdomf but I am confused now. There was conversation earlier in this thread that stated that the wisdom file had to be calculated for the version of fftw you were using. In looking at the default installation, it appears that I have NO fftw packages installed by default.

ETA: Nevermind my original question - based on this post (
https://einsteinathome.org/node/196560&postid=138561, it appears that the 3.3.2 is the version I need to be using on the PI2. I just rebooted the board after adding it - we'll see what happens!

(Old dumb question)
So is my wisdom file optimized for some hidden component within the EAH binaries? Do I need to install the FFTW packages in order to get benefit from wisdom? Do I have a useless wisdom file in place?

My YouTube Channel: https://www.youtube.com/user/KF7IJZ
Follow me on Twitter: https://twitter.com/KF7IJZ

Bikeman (Heinz-Bernd Eggenstein)
Bikeman (Heinz-...
Moderator
Joined: 28 Aug 06
Posts: 3522
Credit: 728145867
RAC: 1194736

RE: So is my wisdom file

Quote:

So is my wisdom file optimized for some hidden component within the EAH binaries? Do I need to install the FFTW packages in order to get benefit from wisdom? Do I have a useless wisdom file in place?

The "hidden component" is the FFTW library. There are two ways to use libraries: statically and dynamically bound libraries.

The E@H apps have libraries like FFTW (which might or might not be installed on volunteers' computers) statically bound to the code, which just means the code of the library is included in the executable that is downloaded itself. For the volunteers this means that they don't need to install any such libraries for an app to run, and for the project's scientists this means they can be sure which version of a given library is used by the app (the one included). Otherwise different volunteers would use different versions which is problematic for testing the app, validating results and reviewing the application.

HB

KF7IJZ
KF7IJZ
Joined: 27 Feb 15
Posts: 110
Credit: 6108311
RAC: 0

RE: The "hidden

Quote:

The "hidden component" is the FFTW library. There are two ways to use libraries: statically and dynamically bound libraries.
*snip*
HB

Thanks for taking the time to explain in detail. This made sense and after rebooting, I immediately saw an improvement. Now to let it run and get the RAC up!

Also, still trying to get my 4 node Pi Zero cluster up and running, but I keep encountering networking issues due to my ignorance of TCP/IP!

My YouTube Channel: https://www.youtube.com/user/KF7IJZ
Follow me on Twitter: https://twitter.com/KF7IJZ

KF7IJZ
KF7IJZ
Joined: 27 Feb 15
Posts: 110
Credit: 6108311
RAC: 0

Just a heads up... After

Just a heads up...

After doing an apt-get upgrade last night, my nodes can no longer contact the EAH servers due to certificate errors. ca-certificates was upgraded as well as ssl components. I am running Jessie Stable.

Follow along in this thread: https://einsteinathome.org/node/198389

My YouTube Channel: https://www.youtube.com/user/KF7IJZ
Follow me on Twitter: https://twitter.com/KF7IJZ

Comment viewing options

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