Re: [fluid-dev] Multi-kernel system test

[David Henningsson]

On 2010-09-30 13:23, Bernd Casper wrote:
> Hi David,
> this is quite interesting.

It is.

> In a real-time scenario using two kernels *slows down* latency and
> overall performance here, when using many programs simultaneously, while
> seemingly increases polyphony performance at te same moment (personal
> impression of my ears - stereo horizon is enlarged). I use
> buffers/size/polyphony 6/384/1024, and 8 instances of FS.
> Is this what's intended?

It is not intended to use more than one instance of FS. This means more 
than one reverb engine, which in turn would consume more CPU in total.

So in theory, for best performance, I would recommend you to

1) use only one instance of FS (and increase polyphony 8 times instead, 
to compensate)

2) use 2 buffers only - and increase size to 1024 to compensate. You 
then might be able to decrease size to "trim" your system for better 
latency.

3) experiment to see if increasing the "kernels" parameter, gives better 
or worse performance. Extending it beyond the number of physical CPU 
cores should give worse performance.

So if this gives the best performance in practice is more than I know - 
 I'm not a computer scheduling expert, especially not on Windows.

(Btw, a few weeks ago I tried to set up jOrgan on Ubuntu Maverick here, 
but I never got any sound output. I had to use jdk though, since 
sun-java wasn't available for Maverick. Perhaps you or Sven could help 
me out some day.)

// David