On Thu, 2003-09-25 at 11:16, Nathan Straz wrote:
> On Thu, Sep 25, 2003 at 05:20:23PM +0200, Christoph Klocker wrote:
> > then I did the benchmark how they were described in the man page I found on
> > the web.
> > [root@localhost lmdd]# ./lmdd if=internal of=/raid/XXX count=1000 fsync=1
> > 8.1920 MB in 2.2116 secs, 3.7041 MB/sec
> >
> > unmounted and mounted again
> > [root@localhost lmdd]# ./lmdd if=/raid/XXX of=internal
> > 8.1920 MB in 0.0776 secs, 105.6065 MB/sec
> >
> > the 105 MB/s seem to be very slow,
>
> I find that you really have to increase the block size you're using for
> I/O. Try adding "bs=1024" or similar to your command lines. Also try
> "odirect=1."
bs=128k or some other larger value might be more meaningful. It looks
like it is defaulting to 8K I/O buffers. A file size of 8M is going
to be completely in cache when you read it back, so you are looking
more at memory speed than anything else here. lmdd is part of lmbench,
this contains tools to measure memory bandwidth too, if your memory
bandwidth is going somewhere else, you need to do some investigation.
Just using a simple scsi disk:
lmdd if=internal of=fred count=1000 fsync=1
8.1920 MB in 0.4958 secs, 16.5234 MB/sec
lmdd if=fred of=internal
8.1920 MB in 0.0355 secs, 230.9753 MB/sec
adding O_DIRECT:
lmdd if=fred of=internal direct=1
8.1920 MB in 0.4241 secs, 19.3180 MB/sec
and on output, demonstrating differing buffer sizes with direct I/O
lmdd if=internal of=fred count=1000 fsync=1 direct=1
8.1920 MB in 2.4649 secs, 3.3234 MB/sec
lmdd if=internal of=fred count=1000 bs=32k fsync=1 direct=1
32.7680 MB in 3.6736 secs, 8.9200 MB/sec
lmdd if=internal of=fred count=1000 bs=128k fsync=1 direct=1
131.0720 MB in 7.7335 secs, 16.9487 MB/sec
lmdd if=internal of=fred count=1000 bs=256k fsync=1 direct=1
262.1440 MB in 14.2395 secs, 18.4096 MB/sec
Steve
--
Steve Lord voice: +1-651-683-3511
Principal Engineer, Filesystem Software email: lord@xxxxxxx
|