On Sun, Dec 1, 2013 at 11:49 AM, Fons Adriaensen <fons@linuxaudio.org> wrote:

On Sun, Dec 01, 2013 at 09:08:37AM -0500, Paul Davis wrote:

> in ardour, the purpose of the disk i/o thread (the "butler") is
> primarily to organize buffering within user space based on observed (but
> very occasional) delays in I/O. it is not attempting to do better than the
> kernel in handling read-ahead and in fact for quite a lot of situations,
> ardour really does rely on the kernel's own buffering to do a good job.

But surely ardour does look ahead in the timeline, to open()/seek()/read()
a new file a few seconds before the data is actually required ? Or does
it keep all files open all the time ?

basically, all files are kept open all the time. in reality, because some systems impose limits on the number of open files, we keep a LRU cache for file descriptors. most users never will run into a situation where the cache is actively used. we read ahead by an amount decided by the user (default is 5 seconds). we write in chunks of 256kB because very old experiments suggested that on ext3 we got the best throughput with that (this predated our use of libsndfile, however.

Suppose you have a timeline like ABCACDBABCCA where each char is a
very short region (maybe just a fraction of a second) and the actual
char refers to the file that region is read from. For example in 'CC'
near the end the first C could be a few seconds of music ending in
silence and the second C a repeat of 400 ms of the end of the first
(e.g to remove a cough or the sound of a score page being turned),
while both are just a few seconds away from the earlier occurences
of C. Would each of the regions be handled independent of any others ?

ardour does not attempt to optimize this. it will call read() with the file descriptor corresponding to A 4 times, the one corresponding to B 3 times, the one corresponding to C 4 times and the one corresponding to D once. each time, it will read the same data each time. this is what i mean by relying on the kernel's own buffering. those second reads typically get an estimated bandwidth on the order of GBytes/second (i.e. basically just memcpy from the kernel buffer cache into user space, because the kernel already has the data cached). you could construct arbitrary scenarios where the kernel's buffering strategy is wrong, but i'm not interested in the work involved in designing a better strategy. there were papers i read in the early 2000's from the late 1990's that outlined such designs (and filesystems for audio) - AFAICT, all of this work has basically been made irrelevant by contemporary hardware and OS performance. when you can stream 100 32 bit float mono interleaved tracks from a single (cheap) SATA disk without real issues, i frankly don't see the point in design work to squeeze a few more tracks out of the same setup.