Date: Mon, 23 Apr 2001 13:45:54 +0200 From: Niek Bergboer <niek@wit379119.student.utwente.nl> To: Alfred Perlstein <bright@wintelcom.net> Cc: freebsd-hackers@freebsd.org Subject: Re: UFS block size vs. write speed Message-ID: <20010423134554.A57241@wit379119.student.utwente.nl> In-Reply-To: <20010420055426.Q1790@fw.wintelcom.net>; from bright@wintelcom.net on Fri, Apr 20, 2001 at 05:54:26AM -0700 References: <20010420144543.F30241@wit379119.student.utwente.nl> <20010420055426.Q1790@fw.wintelcom.net>
next in thread | previous in thread | raw e-mail | index | archive | help
On Fri, Apr 20, 2001 at 05:54:26AM -0700, Alfred Perlstein wrote:
> > PS: The tests were already done with the fs mounted async. The
> > drive in question communicates at UDMA/33 on a PIIX4 controller in
> > an AMD K6/2 233 system.
> It's funny, but you have the ideal system for an interesting
> optimization I've always wanted to try. Since you seem to be
> reading over the network, have you tried doing this, creating
> the file and then using ftruncate on it to extend it, then use
> mmap() and read directly from the socket into the mmap'd area.
I've implemented a quick hack on the BSD ftp-client: in the original
recv-file function data is read from a socket into a buffer, which is
then written to a file. I've mmap-ed the file, and rather than reading
from the socket into the buffer, I read directly from the socket into
the mmaped region. I use the MAP_SHARED and MAP_NOSYNC flags, and
especially the latter makes a huge difference.
The files are read using FTP with a buffer size of 64 kb. They are
read from a Linux 2.4 machine with cartloads of memory and I make sure
the Linux box has the files cached in memory. The Linux machine is
connected to the same 100 MBit/FDX switch and network congestion is
virtually non-existant.
Because the FreeBSD machine (the K6/2 233) only has 64 MB of RAM,
I've made the files small (10 files of 4 MB each, and later 5 files of
8 MB each).
The normal BSD ftp-client reads all the files a approx. 6.5 to 7
Mbyte/s and this speed is more or less constant for all the files. The
mmap-ed version of the ftp-client reads the first few files at 8 to
8.5 MByte/s after which the throughput collapses due to the limited
amount of memory available for caching. I've noticed that using
madvise() with MADV_WILLNEED or MADV_SEQUENTIAL does not make that
much of a difference.
Conclusion: using the mmap trick, the box displays the ugly "all your
system is one big write cache" Linux-behaviour ;) Until memory is full
that is, after which throutput becomes lousy (4.5 Mbyte/s).
> You may have to experiment with several different madvise() flags
> to get optimal performance. Or you may discover that doing this
> "trick" actually makes performance worse because of the way
> the trick screws with what the vm system expects.
> I think a combination of MADV_SEQUENTIAL and/or MADV_WILLNEED
> could do the trick.
See above.
Niek
--
Conscience doth make cowards of us all.
-- Shakespeare
To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-hackers" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20010423134554.A57241>