Advice Wanted on File Servers for a Diskless Network

[Karl Kleinpaste]

duncan at comp.vuw.ac.nz writes:
   Our department will shortly be buying a small number of (20 or so)
   diskless workstations for staff and graduate student use.

   The options currently appear to be [Sun3/[56]0's or HP 9000/300's].

   The two problems are deciding on the appropriate file servers, and
   figuring out which workstation(s) to buy.

Personally, I would recommend the Suns.  We're using both here, and my
clear preference is for the Suns.  The HPs have enough differences in
them that they cause quite a bit of grief in convincing them to cope
with the heavily BSD-and-NFS environment in which we work.  The HPs
are considerably more SysVish than BSDish, with the result that a lot
of things on them are incompatible at some level with our BSD systems,
notably including our Pyramid 98x, on which reside major things like
the department-wide /usr/spool/mail filesystem.  Consider the problems
of a SysV mailer, believing in setgid-mail delivery, as opposed to a
BSD mailer, using setuid-root delivery.  We found a solution to this,
so that the HPs can mount Tut:/usr/spool/mail like everybody else, but
let's just say that comedy is not pretty.

There is, of course, the 14-char filesystem.  Also, HPs do funny
things with a single filesystem which is shared among the server and
clients, with personalizations for each workstation being accomplished
with something HP calls a `CDF,' a Context-Dependent File; a CDF is
actually a directory which contains filenames which are the
workstation names, and in turn those files contain the contents of the
intended filename.  If you look at /etc/passwd, it looks different on
each system, but if you cd /etc/passwd+ (the `+' is how you talk about
a CDF in its directory form), you can see machine-name filenames.
It's weird stuff, and if you're mostly a BSD shop as we are, these
sorts of things can be quite a problem.

   The local Sun agents have suggested that for 20 or so workstations
   we might be looking at two Sun 3/180's to handle the load.

That's close, but they might be trying to sell you too much hardware.
We have 11 Sun3/180's; we put 12-20 clients apiece on them.  You'll
find that ND won't serve more than 20 clients (it's a hard limit), so
if you're going to >20 workstations, >1 3/180 is an absolute
necessity.  Performance is quite acceptable if you have sufficient
disc controllers.  That is, one of our 3/180's is currently serving
two Super Eagles off of a single Xylogics 450 (or is it a 451? can't
remember now) controller.  This is also the server with 20 clients -
paging activity can cause real problems.  We're anxiously awaiting the
re-installation of the 2nd controller there.  On any system with 1
controller per disc, things are really fine and acceptable.  The
problem is in paging activity.  Buy workstations with LOTS of memory,
to cut down your paging traffic.

   HP's
   suggestion was similar but based on HP 9000 model 350's.

I can't say much about load on those systems, other than to say that
we have 1 server with 6 clients, and it has certain perf. problems.
It is not clear at this time if it is due to hardware infant mortality
or software problems.  We know that we are getting some long packets
on our backbone ethernet (1524 bytes), and that in turn is causing us
trouble with our Proteon gateway boxes that connects our dept with the
rest of the OSU campus network.

   Another option is that as part of the same equipment proposal we
   hope to upgrade our current pyramid 90x (8Mb main memory - no data
   cache) to a pyramid 9805 or 9810 with probably 16Mb memory).  The
   pyramid's main task would be to support 30-50 undergraduate and
   graduate student's (not simultaneously).  Would it also be able to
   support the load of acting as a file server for the above number of
   diskless workstations doing typical (whatever that is) work?

The Pyramid can handle it after upgrade.  I wouldn't try it before
upgrading, though, especially without the cache.  Our 98x has 32Mb and
serves 30-50 people simultaneously during the heavier part of the work
day.  It's currently mid-morning, and we are already getting
substantially loaded; by 2pm or so, there'll be 45 people logged in.

[111] [9:19am] tut:/dino0/karl> uptime
 9:19am  up 2 days, 23 hrs,  30 users,  load average: 4.11, 3.40, 3.05
[112] [9:19am] tut:/dino0/karl> 

He's got a fair amount of disc on him, 6 Eagles, which would be 7 if
the 7th were working.

/dev/iop/pdisk00a      19102   16042    1148    93%    /
/dev/iop/pdisk02a      19262     170   17164     1%    /tmp
/dev/iop/pdisk00c      47950   40946    2208    95%    /usr
/dev/iop/pdisk10a      19318    8882    8504    51%    /usr/spool
/dev/iop/pdisk00g     290054  256612    4436    98%    /u0
/dev/iop/pdisk01h     387318  332188   16398    95%    /u1
/dev/iop/pdisk10f     338822  285804   19134    94%    /u2
/dev/iop/pdisk11f     338558  310332   28226    92%    /u3
/dev/iop/pdisk03i     367646   29154  338492     8%    /usr/spool/news
/dev/iop/pdisk03a      19262    4068   15194    21%    /usr/lib/news
/dev/iop/pdisk02b      29054      18   26130     0%    /ai0/sys
/dev/iop/pdisk02f     338558  260712   43990    86%    /ai0

Tut also mounts 14 filesystems from assorted Sun servers and other
Pyramids (we also have 2 98xe's and 2 90x's).  Most of those
filesystems are exported to 140 Suns, about half of which are
inhabited by grad students, faculty, and staff, and the remaining half
by undergrads using filesystems resident on systems other than Tut.
Tut slows down in mid-afternoon when his direct timeshare load hits 50
users and the workstation load is similarly high, but he keeps working
and things keep getting accomplished.  He's configured with about
110Mb of swap space (3 Eagle `b' partitions and one `a').

   Regarding the choice of workstation.  I believe it is possible for
   Sun workstations to boot over NFS now, rather than using the old
   "nd" protocol.  Would this be possible with a Pyramid acting as a
   file server, or would we need at least one Sun file server?

Pyramid demo'd discless Sun service at the Dallas UniForum in
February.  It's not ready for release, apparently (they were using a
hacked SunOS 3.2, from what I was told), but it's Almost There.

   Can
   the HP workstations running HP/UX also boot over NFS?  Would they
   be able to boot from a non-HP file server?  How well could we
   expect a mixed environment of diskless Sun's and HP's (and possibly
   PS/2's) to work?

Sorry, don't know, such ideas don't fit in with our plans/goals.

Hope this helps,
--Karl