Version française
Home     About     Download     Resources     Contact us    
Browse thread
How to read different ints from a Bigarray?
[ Home ] [ Index: by date | by threads ]
[ Search: ]

[ Message by date: previous | next ] [ Message in thread: previous | next ] [ Thread: previous | next ]
Date: -- (:)
From: Gerd Stolpmann <gerd@g...>
Subject: Re: [Caml-list] Re: How to read different ints from a Bigarray?

Am Freitag, den 30.10.2009, 00:43 +0100 schrieb Goswin von Brederlow:
> Gerd Stolpmann <gerd@gerd-stolpmann.de> writes:
> 
> > Am Donnerstag, den 29.10.2009, 21:40 +0100 schrieb Florian Weimer:
> >> * Goswin von Brederlow:
> >> 
> >> > - The data is passed to libaio and needs to be kept alive and unmoved
> >> >   as long as libaio knows it.
> >> 
> >> It also has to be aligned to a 512-byte boundary, so you can use
> >> O_DIRECT.  Linux does not support truely asynchronous I/O without
> >> O_DIRECT AFAIK, which rarely makes it worth the trouble.
> >
> > Right. There is also the question whether aio for regular files (i.e.
> > files backed by page cache) is continued to be supported at all - it is
> > well known that Linus Torvalds doesn't like it. It can happen that at
> > some day aio will be restricted to block devices only.
> >
> > So I wouldn't use it for production code, but it is of course still an
> > interesting interface.
> >
> > Gerd
> 
> Damn. That seems so stupid. Then writing asynchronous will only be
> possible with creating a pot full of worker thread, each one writing
> one chunk. So you get all those chunks in random order submitted to
> the kernel, the kernel has to reorder them, fit them back together,
> write them and then wake up the right thread for each piece
> completed. So much extra work while libaio has all the data already in
> perfect structures for the kernel.

Well, this is exactly the implementation of the POSIX aio functions in
glibc. They are mapped to a bunch of threads.

> And how will you do barriers when writing with threads? Wait for all
> threads to complete every time you hit a barrier and thereby stalling
> the pipeline?

You can't implement barriers. When you have page-cache backed I/O (i.e.
non-direct I/O, no matter of aio or sync I/O) there is no control when
data is written. Ok, there is fsync but this is very coarse-grained
control.

Gerd
-- 
------------------------------------------------------------
Gerd Stolpmann, Bad Nauheimer Str.3, 64289 Darmstadt,Germany 
gerd@gerd-stolpmann.de          http://www.gerd-stolpmann.de
Phone: +49-6151-153855                  Fax: +49-6151-997714
------------------------------------------------------------