Re: [Discuss-gnuradio] Thread scheduling for high-sensitivity, high-resolution FFTs

[Eric Blossom]

On Tue, Jan 06, 2009 at 06:58:47PM -0500, Marcus D. Leech wrote:
> Let's say, for the sake of interesting discussion, that I have access to
> a beastly multi-CPU 'puter, perhaps based on the
>   latest Xeon 7460 series CPUs, and a quad-socket server, giving me 24
> CPUs, and 48 threads.
> 
> I'm interested in wideband, high-sensitivity, high-resolution spectroscopy.
> 
> I envisage a scheme where one thread gathers an FFT-frame worth of data
> (let's say from an USRP2), and doles it out to one
>   of a number of other threads that are doing FFTs on these frames of
> data.  All the outputs of the FFT threads are averaged
>   together to produce a result that is then time averaged.
> 
> The existing FFT display schemes in Gnu Radio use a "keep-one-in-n"
> scheme for the FFTs, which means that sensitivity
>   is necessarily lost, due to the spectrum being under-sampled in the
> time domain.
> 
> Is there existing plumbing to be able to try this sort of thing? [On a
> smaller scale--I don't actually have such a monster server
>   at my disposal at the moment!!].
> 

Is this an SMP machine, or a cluster?

If SMP, it'll "just work".  If it's a cluster, you'll need to figure
out how to partition separate instantiations of GR on each node.

Eric