From owner-freebsd-current  Sat Apr  8 04:49:35 1995
Return-Path: current-owner
Received: (from majordom@localhost)
          by freefall.cdrom.com (8.6.10/8.6.6) id EAA13216
          for current-outgoing; Sat, 8 Apr 1995 04:49:35 -0700
Received: from godzilla.zeta.org.au (godzilla.zeta.org.au [203.2.228.34])
          by freefall.cdrom.com (8.6.10/8.6.6) with ESMTP id EAA13193
          for <freebsd-current@freebsd.org>; Sat, 8 Apr 1995 04:48:20 -0700
Received: (from bde@localhost) by godzilla.zeta.org.au (8.6.9/8.6.9) id VAA26041; Sat, 8 Apr 1995 21:44:23 +1000
Date: Sat, 8 Apr 1995 21:44:23 +1000
From: Bruce Evans <bde@zeta.org.au>
Message-Id: <199504081144.VAA26041@godzilla.zeta.org.au>
To: phk@ref.tfs.com, taob@gate.sinica.edu.tw
Subject: Re: Disk performance
Cc: freebsd-current@FreeBSD.org
Sender: current-owner@FreeBSD.org
Precedence: bulk

>Just to remind you how far we have got:

>Seagate ST506, 20Mb "BIOS Type 2" disk:
>	dd if=/dev/rwd0d of=/dev/null bs=8k count=200
>	200+0 records in
>	200+0 records out
>	1638400 bytes transferred in 10 secs (163840 bytes/sec)

>Imprimis(?) 94186-383 ESDI disk:
>	dd if=/dev/rwd2d of=/dev/null bs=8k count=400
>	400+0 records in
>	400+0 records out
>	3276800 bytes transferred in 8 secs (409600 bytes/sec)

Just to remind you that IDE is better than SCSI ;-):

[All with profiling kernel, profiling off]
Toshiba MK537FB disk 486DX2/66 VLB Ultrastor 34F controller:
	dd if=/dev/rsd0 of=/dev/null bs=512 count=4000
	4000+0 records in
	4000+0 records out
	2048000 bytes transferred in 20 secs (102400 bytes/sec)
	dd if=/dev/rsd0 of=/dev/null bs=4k count=4000
	4000+0 records in
	4000+0 records out
	16384000 bytes transferred in 21 secs (780190 bytes/sec)
	dd if=/dev/rsd0 of=/dev/null bs=8k count=4000
	4000+0 records in
	4000+0 records out
	32768000 bytes transferred in 26 secs (1260307 bytes/sec)
	dd if=/dev/rsd0 of=/dev/null bs=64k count=4000
	4000+0 records in
	4000+0 records out
	262144000 bytes transferred in 102 secs (2570079 bytes/sec)

Samsung SHD-3212A disk 486DX/33 IDE:
	dd if=/dev/rwd0 of=/dev/null bs=512 count=4000
	4000+0 records in
	4000+0 records out
	2048000 bytes transferred in 4 secs (512000 bytes/sec)
	dd if=/dev/rwd0 of=/dev/null bs=4k count=4000
	4000+0 records in
	4000+0 records out
	16384000 bytes transferred in 13 secs (1260307 bytes/sec)
	dd if=/dev/rwd0 of=/dev/null bs=8k count=4000
	4000+0 records in
	4000+0 records out
	32768000 bytes transferred in 25 secs (1310720 bytes/sec)
	dd if=/dev/rwd0 of=/dev/null bs=64k count=4000
	4000+0 records in
	4000+0 records out
	262144000 bytes transferred in 188 secs (1394382 bytes/sec)

The SCSI disk has a much lower overhead but it has a worse transfer
rate for a block size of 8K, and even worse transfer rate for smaller
block sizes.  It only has a reasonable transfer rate for huge block
sizes (64K or more).  Such block sizes get used when iozone is run
but don't get used much for normal operations.  Operations such as
`cvs co' that do a lot of small i/o's run much faster on the IDE
drive.

A high-end SCSI disk is competitive with the low-end IDE disk for
a block size of 4K:

Quantum XP34301 disk 486DX2/66 VLB Ultrastor controller:
	dd if=/dev/rsd1 of=/dev/null bs=512 count=4000
	4000+0 records in
	4000+0 records out
	2048000 bytes transferred in 12 secs (170666 bytes/sec)
	dd if=/dev/rsd1 of=/dev/null bs=4k count=4000
	4000+0 records in
	4000+0 records out
	16384000 bytes transferred in 14 secs (1170285 bytes/sec)
	dd if=/dev/rsd1 of=/dev/null bs=8k count=4000
	4000+0 records in
	4000+0 records out
	32768000 bytes transferred in 19 secs (1724631 bytes/sec)
	dd if=/dev/rsd1 of=/dev/null bs=64k count=4000
	4000+0 records in
	4000+0 records out
	262144000 bytes transferred in 80 secs (3276800 bytes/sec)
					       (5MB/sec before memory is
					        fragmented)

Most of the SCSI slowness seems to be caused by the extremely high
command overhead of the U34F controller and/or SCSI interface.  The
minumum overhead seems to be about:

	Samsung/IDE/noname    on 486DX/33:	1msec
	Quantum/SCSI/BT445C   on 486DX2/66:	1.7msec (dd output not shown)
	Toshiba/SCSI/BT445C   on 486DX2/66:	3msec (dd output not shown)
	Quantum/SCSI/U34F     on 486DX2/66:	3.3msec
	Toshiba/SCSI/U34F     on 486DX2/66:	5msec

Bruce