Subject: Re: SCSI/disklabel questions
To: Michael Wolfson <email@example.com>
From: David Brownlee <firstname.lastname@example.org>
Date: 06/23/1998 22:28:22
On Mon, 22 Jun 1998, Michael Wolfson wrote:
> Hi folks, I've got a few burning questions.
> I recently got my hands on a reasonably fast SCSI-2 drive. On other
> platforms, I can get a sustained transfer rate of well over 3.5 MB/s on
> slow/narrow SCSI. On my 400s, the best I've ever seen is 700 KB/s (actual
> elapsed time, not processor time).
> Where does the slowdown come from: the SCSI hardware chipset in the 400,
> or the BSD drivers (i.e. will performance get better once people devote
> time to writing improved SCSI code)?
My guess is switching to the NetBSD MI scsi driver will certainly
> Next, on the topic of disklabels, I seem to recall a discussion a while
> ago about labeling a drive that doesn't necessarily have perfect geometry
> (or unknown geometry). I don't believe there was any real consensus on the
> topic. Could someone semi-authoritatively comment on the following:
> Use default disklabel or (if known) drive parameters (e.g. from
> www.seagate.com)? Note, I've never gotten the default disklabel to
> be reasonable, the cylinders are always set to 1.
Thats a bug :) - the default faked up disklabel _should_ work for
the disk, and should include a correct 'c' (or if you have i386
drainbamage 'd') partition.
> If not a divisible number of sectors, how to lay out the disk (i.e.
> presume that the last cylinder is only partially filled)?
Given the geometry of any modern disk is not something that can
be reduced to C/H/S, you can either pick something that 'looks
close', or just pick something convenient (such as 1MB cylinders)
> What's the most efficient setup for ccd (i.e. disk striping)? I could
> only get ~600 KB/s out of two 2213A drives, but measured ~610 KB/s
> when they weren't arranged in ccd.
That depends on where the bottleneck is. Can you try timing two
simultaneous 'dd's from the disks' raw partitions, vs the same
dd on one drive. The driver & interface could be the bottleneck
in which case all ccd can gain you is larger partitions.
> Does geometry *really* affect performance on slower workstations/drives
> (e.g. a HP-IB disk on a 320)?
The geometry optimisations were designed around disks of that era
:) (Such as great clunking RD54 and RD81's on vaxes :) No cache
and a seek time that just terrifies modern thinking. Geometry
optimisations can really help :) As soon as you put a fair sized
cache, and use variable sectors per track, it all becomes pretty
> Is sector 0 in the disklabel the inside of the disk or outside, and
> wouldn't it make sense to put swap on the inside for best performance?
Uh.. inside (on old drives). Modern drives are free to do whatever
they feel like, though I'd expect that convention to be adhered
There are two things that affect performance here - the speed at
which data can be read off the disk, and (much bigger) how long it
took the head to get to where it needs to be.
To minimise seek time put swap in the middle of the disk.
For variable sectors per track its better to put the swap towards
the outside where there are more sectors per track (the disk
always spins at the same speed, unless you have a floppy drive
designed by an apple engineer on strange drugs... dont go there)
> How are the bootblocks and disklabel arranged? As best as I can
> determine, the label sits at sector 1, and the boot blocks surround it
> at sector 0 and 2 thru whatever. This would make sense since improperly
> installed boot blocks can be wiped out by disklabel and vice versa.
Doesn't the hp300 port have some bizarre LIF filesystem at the
> The install notes recommend ofsetting the a partition by one cylinder,
> which is fine for disks with small cylinder sizes, but how much space
> is really necessary for SCSI disks?
Uh... pass - just make your cylinders a convenient size :)
(When you find out what size that is)
> What other NetBSD platforms can/can't write disklabel and bootblocks for
> hp300 machines (I know sun3x, sun3, and sparc can't)?
Its probably the LIF bootblock filesystem. I'm sure they could if
someone tweaked the code.
> On a slightly different track: what SCSI DAT drives are supported, and how
> does one go about getting more models supported? I have a HP C1504C
> DAT (a.k.a. 35480A) from a 730 that's not recognized (and thus not
I don't even have a working audio tape drive with me at the
moment... pass :)
> And, lastly, can anyone comment on why the HP-IB isn't available when one
> or more SCSI devices are attached (i.e. can't use HP-IB drives or
> tapes if a SCSI drive is plugged into my 400s)? This used to work
> thru NetBSD-1.1.
... and I'm more lost here - is there a PR open on it?
-=- "Promise carved in stone, deeper than the sea" -=-