Summary of Responses to Disk Partitions on large disks

[brown at kpno.UUCP]

This is the much overdo summary of disk layout replies I promised over a
month and a half ago.  Some of the comments are appropriate to all versions of
Unix I am familiar with, others don't apply to the 4.* bsd systems from
Berkeley.

    To everyone who responded to my question, Thank you very much and
I apologize for the delay in putting this summary together.
--------

1.	Go to 6250 bpi tape drives

2.	Individual small file systems are a pain because individual
	partitions run out of space.
	Also on large drives you can easily have more than 8 partitions
	if the drive is large (If you didn't know, 8 is the maximum number
	of device partitions. 40Mb = 1 tape level 0 dump, so 320 Mb disk max)

3.	Set the drive up with a clear partition for "dd" backups.
	(dump is better for non-full drives, dd for full)

4.	Move root (/) to the  center of the disk (fix /boot)
	(Interesting but not really what I was after)

5.	V7 systems mung file systems easier than Berkeley so keep
	the partitions small to minimize info lossage.

6.	divvy the disk up in partitions ~33 Mb so a level 0 dump
	will fit comfortably on a single 2400' tape.  You can
	stack many level 9 dumps on a single tape, though this
	makes the restores a bit tricky.

7.	don't do level 0 dumps as often.
		level 0		not more often than 6 mos.
		level 3		each month
		level 5		each week
		level 9		daily
		If the dailys get too big add a level 7 midweek.

    My own biased opinions after reviewing these responses and considering
my situation are:
	1  I'd love a 6250 tape drive, there seem to be cheaper ones now.
	2,3,5 & 6
		I think these are not appropriate for large disks, see caveat
		on 2 about partitions.
	4	I'm gonna try this on our RA81s if we ever get 4.2BSD
		and our new 750.
	7	Sounds like a better schedule than our current one.

One person responded with a nice discussion of the whole situation of
disks on Unix and some goals to keep in mind while configuring the system.
I'll just reproduce his letter below for you:
============================================================================

Subject: Re:  Disk partition sizes on big disks

My question is, if you have four smaller partitions, each of which takes an
entire tape to backup, have you gained anything?

There are several things to consider in allocating disk partitions.

1) You can only specify eight distinct partitions per controller-type.
2) Certain system-related file systems get hit harder than others.
3) Load should be balanced over all seek arms.
4) /tmp should be a mounted file system.
5) You should swap off multiple drives.
6) Static file system pieces should be isolated.
7) The Bell or Berkeley-distributed partitions are not sacred.

Some explanation.  Most people consider it good practice to have root, /usr,
and /tmp on separate disks whenever possible.  Next, spread users evenly
over the drives.  This really makes a noticable difference.

If your system crashes, chances are /tmp will have file system inconsistencies,
since, in theory, all files in /tmp are open most of the time.  I've found it
useful to make /tmp a separate file system and put a mkfs or newfs for it in
my /etc/rc before the mount -a.  Then, naturally, you don't fsck it.

Another useful thing to do is mount /usr/spool.  The result is that
/usr then becomes a relatively static file system and can be dumped
less often.  We have found that incremental dumps of the remaining
/usr comprise mostly log files from /usr/adm (like wtmp and acct).

There is a motivation for mounting /usr/src on a disk other than the 
one that /usr is on - if you lose /usr due to serious failure, you
can reconstruct most of it from sources.  This also lets you mount
/usr/src read-only on those machines that simply hold viewable (not
modifiable) sources.

On RA81's, the Berkeley-distributed partitions are a mess and should
be completely revised.  I chose partitions in even cylinder group
multiples (i.e. multiples of 16 cylinders).  For example, my 'a'
partition is 32 cylinders (22848 blocks) and 'b' is 48 cylinders
(34272 blocks).  For the remaining space, I have one partition that
covers it all, two more that divide that roughly in half, and two more
that divide the first of those in half (about 30/70).  If this sounds
vaguely like a buddy system, you're right.  I kept 'c' has the full-disk
partitions, which is basically useless in winchesters.

If you're not running 4.1c or 4.2 BSD, having a very large partition
is not such a good idea, since all the inodes are  concentrated at the
beginning of the file system.

One more detail.  On RA81s, or any disk for that matter, make sure
your partition size mod your block size is zero.  Fsck fails otherwise
when it tries to read the 'last' block on the partition.

(Name Deleted)
(Institution Deleted)

PS Get a 6250 BPI tape drive if you possibly can.

=============================================================================

	regards,
	Mike Brown	Kitt Peak National Observatory
			Tucson, Arizona			(602) 325-9249	

UUCP:	{...,allegra,arizona,decvax,hao,ihnp4,sdcsvax,unc} !kpno!brown
ARPA:	kpno!brown at LBL-CSAM.ARPA