Subject: Re: [linux-audio-dev] Re: costs of IPC
From: Steve Harris (S.W.Harris_AT_ecs.soton.ac.uk)
Date: Wed May 16 2001 - 00:00:44 EEST
On Tue, May 15, 2001 at 03:53:46PM -0400, Paul Davis wrote:
> >Thats all very well, but I suspect that a lot of apps will needa lot more
> >memory than that, certainly anything doing convolution, or a software
> >sampler. If a process goes and mallocs a few dozen meg to keep samples in,
> >what will happen? How does heap size affect context switch time?
>
> heap size by itself isn't an issue, all that matters is how much
> memory is actually touched during each execution of the process
> between context switches.
Yes, I just found that out the hard way ;) Sorry, its late, please ignore
my ramblings in the last mail.
> and its not as bad as you think, since programs like software samplers
> will likely be feeding audio data from disk into a ringbuffer. even
> though the ringbuffer may be large, each invocation by the engine will
> only touch a fairly small amount of it.
Mmmm, yes. I was wondering about that. It's abviously bad for the DSP
parts to do the loading form disk, so what is going to, and how is it
going to feed that data to the DSP part?
> i don't know enough about convolution and other DSP algorithms to
> sense their memory footprint. do you have any thoughts?
Not loads, probably a meg or two, but half of it will get polluted on each
invocation. With this model you could save quite a bit as long as you
didn't mind using the same reverb impulse on every channel.
- Steve
This archive was generated by hypermail 2b28 : Tue May 15 2001 - 23:18:17 EEST