Subject: disk i/o performance
To: None <port-vax@netbsd.org>
From: Chuck Silvers <chuq@chuq.com>
List: port-vax
Date: 06/06/2001 23:39:46
I was playing with filesystem performance on my vax 4000/90 tonight
and found that file i/o was pretty sucky compared to raw:

1 vax:~ # ./pt -f -r -c 20 /dev/rsd0a
20971520 bytes transferred in 8.479 secs (2473074 bytes/sec)

2 vax:~ # ./pt -f -w -c 20 /dev/rsd0a
20971520 bytes transferred in 13.455 secs (1558591 bytes/sec)


3 vax:~ # ./pt -f -r -c 20 /mnt/file 
20971520 bytes transferred in 10.452 secs (2006278 bytes/sec)

6 vax:~ # ./pt -f -w -c 20 /mnt/file
20971520 bytes transferred in 26.582 secs (788925 bytes/sec)



chatting a bit with the icb folks, we noticed that MAXBSIZE was only 16k,
so I tried it again after letting MAXBSIZE default to MAXPHYS (63k).
the results were improved, but writes were still very slow:

4 vax:~ # ./pt -f -r -c 20 /mnt/file
20971520 bytes transferred in 8.584 secs (2442953 bytes/sec)

7 vax:~ # ./pt -f -w -c 20 /mnt/file
20971520 bytes transferred in 21.659 secs (968236 bytes/sec)


increasing MAXPHYS, MAXBSIZE and the asc maxxfer to 64k improved the write
result to where it should be:

4 vax:~ # ./pt -f -r -c 20 /mnt/file
20971520 bytes transferred in 8.583 secs (2443159 bytes/sec)

5 vax:~ # ./pt -f -w -c 20 /mnt/file
20971520 bytes transferred in 13.064 secs (1605254 bytes/sec)



so.  it'd be real nice if we could use the 64k size.
matt thomas was concerned that not all devices could support 64k xfers.
does anyone know if this is correct or not?  if all devices do not
handle 64k xfers, what would it take to make it so that they do?

-Chuck