10 Apr
2006
10 Apr
'06
6:58 p.m.
Paul, Everyone,
What kind of assumptions can we make about the performance of the
various FIFO pipe/socket/whatever mechanisms on a POSIX machine? This
is by no means my area of expertise, but I currently hold the
assumption that a POSIX machine should provide FIFO pipe-ish things
that are *very* fast, considering they dont live in a local context.
If native FIFO I/O mechanisms are not fast enough for, say, spewing
1/256th of a 10-20 second wave form's data to some app, isn't there
something wrong? What metrics do you keep on the tip of your tongue
after working with ardour and jack?
On 4/7/06, Paul Davis <paul(a)linuxaudiosystems.com> wrote:
On Fri, 2006-04-07 at 14:23 -0800, Patrick Stinson
wrote:
--
Patrick Kidd Stinson
http://www.patrickkidd.com/
http://pkaudio.sourceforge.net/
http://pksampler.sourceforge.net/
As an added note to my previous comments, I
really like the app
interface that mpd uses. Writing ascii events using some spec or
another to a file descriptor (socket in mpd's case) seems to be a
terrific way to communicated with apps and libs. I run the svn
pksampler like this:
pksampler | pkaudiod
I could just as easily have either use a socket or some other pipe,
and this is easy to debug, script, profile, etc. On the client and
server sides, I wrap my 10-minute rpc into an interface with generic
calls. Everything having to do with input (MIDI, OSC, whatever)
happens in the app code. The interface functions resemble those of
FMODEx's interface; add sample, connect to channel, set sample attrs,
etc.
As an interface designer, the first thing I look for on an engine's
project site is some sort of asynchronous API - I should never concern
myself with anything outside referencing the api from my app's one
windowing thread. FMOD, gstreamer, and my dead pkaudio project do this
very well. I don't ever want to worry about what thread it happens in,
what threads it will affect, or what the performance effects of
*making* the call will be (as opposed to residual effects).
although i agree that this is the right design for many classes of
application design, i would like to see how you propose to tackle
metering and waveform display (the two most difficult examples).
ardour would be relatively easy to separate into interface+engine
processes (as opposed to just a lib/lib-client separation) if it were
not for these issues. moving waveform and metering data back and forth
between two processes via a wire protocol is very expensive and
inefficient.
--p