This is the mail archive of the
fortran@gcc.gnu.org
mailing list for the GNU Fortran project.
Re: [patch, fortran]PR25829 Add support for F2003 I/O features
- From: Janne Blomqvist <blomqvist dot janne at gmail dot com>
- To: Jerry DeLisle <jvdelisle at verizon dot net>
- Cc: FX Coudert <fxcoudert at gmail dot com>, Fortran List <fortran at gcc dot gnu dot org>
- Date: Tue, 18 Mar 2008 12:04:44 +0200
- Subject: Re: [patch, fortran]PR25829 Add support for F2003 I/O features
- Dkim-signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gmail.com; s=gamma; h=domainkey-signature:received:received:message-id:date:from:user-agent:mime-version:to:cc:subject:references:in-reply-to:content-type:content-transfer-encoding; bh=s2S6af1NfblxxuU7lQ63UraSpSIho2CVjWrfjAHS/ok=; b=OOi3OGloHOqYUc7PMGLVldB5nQTjBGCo7rN+fQgKqChBsfyTPFgc6O8JHckmW/TjFtirJRg7OqHvZJ8G5agkriEHMl4JhR3T85xb3UlnJST4VqGOXhvKPbWS4TlQidfpRMgdh6pRtfn9Od/WI4zU0TKjDXfTn0nDf+4p/SDWleY=
- Domainkey-signature: a=rsa-sha1; c=nofws; d=gmail.com; s=gamma; h=message-id:date:from:user-agent:mime-version:to:cc:subject:references:in-reply-to:content-type:content-transfer-encoding; b=eHEzR+I26MmLIrNZaxu4flIpwU9Wz6U+hF8b9Ggmm22DM7xyYxdqPBn+MGt6F/zM872Hg0F4DG8Zyv64TdAb8eALpO82krDErHO5/rgjtvw4q58alydwxcGB6nHVW03ZG3V+3Wurde6dZ/oX2qNPylpReba8j607ecWyf7O6H4E=
- References: <47DD926B.9090109@verizon.net> <E8862890-0160-4B71-B30E-8E6DB1173CC5@gmail.com> <47DDB5F3.20700@verizon.net> <47DE1495.5000306@gmail.com> <47DF241F.3000504@verizon.net>
Jerry DeLisle wrote:
Janne Blomqvist wrote:
IIRC on Linux AIO requires that the file is opened with O_DIRECT, and
that all I/O is page size (typically 4 KB) aligned. The AIO syscalls
do work if these conditions are not met, but internally they fall back
to a normal synchronous I/O. I think there are some efforts to rectify
this, but I don't think they have been committed to the mainline
kernel yet. Just something to keep in mind when you want to benchmark
the implementation.
Does this mean that we must do fixed size blocks of I/O that are a
multiple of the system page size? or that the memory of our source
buffer must be aligned on a page boundary?
Both of these, I believe. See
http://lwn.net/Articles/148755/
Actually, it seems the requirement is that I/O is block aligned, which
at least on ext3 is typically 4 KB so it's the same as the page size.
I did not read this in the
man pages. Where can I find more info on this.
Some general info about AIO, event completion etc. Mostly about network
I/O, but still a good read:
http://www.kegel.com/c10k.html
Linux AIO:
http://lse.sourceforge.net/io/aio.html
http://www.linuxsymposium.org/proceedings/reprints/Reprint-Bhattacharya-OLS2004.pdf
Linux buffered AIO:
http://lwn.net/Articles/216200/
Threadlets/syslets/fibrils/wtf?:
http://lwn.net/Articles/223432/
General rant about AIO on Linux:
http://davmac.org/davpage/linux/async-io.html
Either way it does make
our "accounting" code a bit more cumbersome.
Yes. OTOH fixing this should probably get rid of some unnecessary
flushes and seeks that we now have as well. :)
Also, for folks who have non-linux systems, what kind of support do they
have for AIO?
I think most Unixes have relatively robust select/poll implementations.
Unfortunately these are not really useful for file I/O. POSIX AIO works
for files, but it is a bit newer so support for that is not as
widespread (and might suffer from various restrictions like the O_DIRECT
+ block alignment on Linux as explained above).
OTOH, on windows NT I understand this works much better, as NT was
designed for AIO from day one (internally in the kernel all I/O is
asynchronous, synchronous I/O is implemented as AIO + wait). The API is
called "I/O Completion Port":
http://en.wikipedia.org/wiki/IOCP
--
Janne Blomqvist