[comp.unix.ultrix] Summary IBM, DEC etc file servers

xxbja@csduts1.lerc.nasa.gov (Betty Jo Armstead) (03/16/91)

Here are the replys I received in response to the post on distributed
servers.
The post was:
>We are in the process of designing an architecture for a centralized mass
>storage system.  The central mass store system will probably include
>some flavor of Unitree and possibly AFS. We think that this design
>will include distributed file servers connected to a fairly high speed
>network.  Therefore, I would like to hear about anyone's experience
>with their favorite file server.  I have already collected some info and
>will be glad to summarize the results.

>From hubler@galaxy.lerc.nasa.gov  Fri Mar  8 10:33:03 1991
Organization: NASA Lewis Research Center, Cleveland, Ohio

We are using an RS6000 in the AAC building and using NFS to export
about 300M of files to the other RS6000's (4) and also a sun and an 
SGI.  You would be welcome to look about and see what we have.  Give
me a call at 3-7014

Dale Hubler


>From se@ikp.Uni-Koeln.DE  Fri Mar  8 15:26:37 1991

We choose the DECstation 5000/200 as a machine for an application were 
we have to deal with lots of rather big files. There are 4 Fujitsu M2266 SCSI 
drives on this machine. This combination gives 2.3 MByte/sec performance 
when reading big Blocks of data (64K/transfer). But with typical file system 
transfers of 8KByte you get at most 1.2 MByte/sec, because of the overhead 
of each SCSI command. With two drives transferring concurrently you get 
800 KByte/sec each (=1.6 MByte/sec max.).
with 1 MByte/sec you get about 20% system time.
We got our machine with a second SCSI controller (the price was ca. $1500).
This controller was intended for slow devices (EXABYTE, CD-ROM,...).
But we may choose to connect two discs to each of the SCSI-controllers 
for better performance.
The ethernet performance is quite good, but I don't have good numbers 
for the system time required. I think, it was about 20% system time at 
950 KByte/sec throughput (what you can expect for FTP on a silent LAN), 
but I didn't try to get exact numbers.
I think the DECstation has good performace at low cost, at least if you
get 3rd party disks/memory.
If you want better performance, I think a Sun with IPI disks is best.
You can expect 3.5 MByte/sec read/write throughput through the 
filesystem with not so expensive parallel head IPI disks on a 4/470.
The DECstation required some tuning to get a speed of 1.1 MByte/sec 
on read/write.
One last point: There is a problem with drives that have a capacity of 
more than 1 GByte. Current ULTRIX device drivers can't address blocks
beyond the first Gigabyte by now ! (Expected to be OK in ULTRIX 4.2)

 Stefan Esser, Institute of Nuclear Physics, University of Cologne, W. Germany
 se@ikp.Uni-Koeln.DE                                            [134.95.192.9]

>From epd74hp!geoff@relay.EU.net  Mon Mar 11 03:07:15 1991

I read your posting on Usenet with interest. We are in a
similar position to yourselves in that we will be setting
up a file server service sometime this year in our office
in the Hague.

Like yourselves, we would like to track down experiences
of others to help in the selection. - The choice will
be between a DECsystem or an IBM RS/6000.

This message is just to urge you to post the results, and
to thank you for (hopefully) starting the ball rolling...

- Geoff Coupe,
  Dept. EPD/74,
  Shell Internationale Petroleum MIJ
  The Hague, Holland
  geoff%epd74hp@nluug.nl


>From @RELAY.CS.NET:pavlov@canisius.edu  Sat Mar  9 23:48:13 1991

  We use a DEC5000 as a file server and as one of our timesharing machines
  for apx. 50 concurrent users peak load (there are additional users, but
  we have a MIPS M/120 as a second file server for convenience).

  This works fine.  But none of the people who use the 5000 file services
  can be considered true "heavy users".

  greg pavlov, fstrf, amherst, ny
  pavlov@stewart.fstrf.org

>From beepy@Eng.Sun.COM  Sat Mar  9 17:19:52 1991
Organization: Sun Microsystems, Mt. View, Ca.

Hi!

I noticed you shotgun posting your "XXX as File Server" question
to several news groups. Did you post a Sun one? (I saw DEC and IBM).

I'd be interested in:

1. Answering questions about using Sun's, or answerinf NFS questions
   in exchange for:

2. A summary of the responses and your "walk away" position given the
   input.

I have on tap a few NFS performance papers from a few sources, and
know another company that has been putting them together. If you're
interested.

Brian Pawlowski
Staff Engineer
NFS Development

P.S. Are you looking at other distributed file systems (other than
     NFS), say like AFS? You didn't say you wanted an "NFS" file server.
     For all I know you want an RS/6000 to act as a Novell file server
     for a bunch of PC's. Could you expand on your application.

     I'm intensely interested in what you conclude from your call
     for input.


>From hudgens@sun13.scri.fsu.edu  Fri Mar  8 23:13:52 1991


We're using a 5000 here as a file server, 32M, two scsi cards, and 
4 1G drives.  It does pretty good, in that it seems to be able to 
keep up with most of the workstations here.  One problem is that 
occasionally it will lock up when a client fills the filesystem
full, or uses a stale filehandle.  In that case, the system will
attempt to spend 99.9% of its time trying to write error messages
to the console, and effectively is locked up.  By staying logged
in as root, and putting the dxsession into a "paused" state,
this can be minimized.  This is probably a feature, but I think 
it makes the fileserver very easily crippled.  

Also, there is no way to boot one of these things with a graphics
card installed but without a monitor.  If you remove the graphics
card and use a serial port, then the above problem becomes 
critical, and worse, the boot process from the "loading vmunix"
to the "foobar login:" is entirely blind, with no messages whatsoever.
So if it hangs for some reason, you must replace the card and monitor
and then power it back up to see what's wrong.  

JHH

Jim Hudgens		Supercomputer Computations Research Institute
hudgens@sun13.scri.fsu.edu

>From morse@mprgate.mpr.ca  Fri Mar  8 20:13:56 1991

We have several DECstation 5000s here, as well as a DECsystem 5000,
which is being used as a file/compute server. There are essentially no
differences between the systems, save for the Turbochannel framebuffer
card. Our server is heavily utilized. We have 4x665 Mbyte drives;
another 1.2 Gbyte drive is on order. Also on it is a TK50 tape drive,
2x8mm tape drives (for backing up the workstations, which have 665
Mbyte drives), and soon a CD-ROM for bookreader, and software
installations. (RIS is installed on it. It's FANTASTIC!!) If you have
been counting, you will note that we ran out of SCCI ports; An
additional Turbochannel SCCI card is also on order. The machine
currently has 40Mbytes of mixed DEC/3rd party memory. I have no idea
how much is normally being used. I suspect not much, though that
depends on how many interactive users are on it. It is a file server
for 2 dozen or so workstations (/usr/local, etc.), and there are
normally a dozen or so interactive users, some of them being DECwrite
users (DECwrite is a memory hog!). The machine has exhibited no
performance problems whatsoever. In short (that wasn't short was
it??), we have no complaints about using a DS5000 as a server. (Though
it is nice to have one all to yourself! 8-)

Daryl Morse                     | Voice : (604) 293-5476
MPR Teltech Ltd. 		| Fax   : (604) 293-5787
8999 Nelson Way, Burnaby, BC    | E-Mail: morse@quark.mpr.ca
Canada, V5A 4B5                 |         quark.mpr.ca!morse@uunet.uu.net

>From lebaron@vicki.inrs-telecom.uquebec.ca  Mon Mar 11 12:27:09 1991

We've got a DECstation 5000 with 32 Megs, 2 RZ57, 1 RZ56, 2 RZ55 on the
SCSI. It's serving 20 2100 & 3100 systems, each with a local swap disk.
Performance is just fine. It does seem to be very sensitive to loads
placed on the server. We've discovered that if somone starts to churn
on the 5000, the clients get very slow very quickly.

The performance is much better than when we were using 2 2100s to serve about
15 clients. The 5000 has much better SCSI performance.

-- 
Gregory LeBaron   lebaron@inrs-telecom.uquebec.ca    (514) 765-7835
INRS Telecommunications, 3 Place du Commerce, Verdun, Quebec, Canada, H3E 1H6


>From bruce@cluster.cs.su.OZ.AU  Wed Mar 13 06:24:12 1991
And I would be interested to hear what you discover, please.
Either by mail or a posting.  What is your centralized mass
storage system?  We are about to do something more modest
for a 20GB juke-box.  Perhaps we can swap war stories!
(What is a "Unitree"?)

Cheers,
bruce.

Bruce Janson
Basser Department of Computer Science
University of Sydney
Sydney, N.S.W., 2006
AUSTRALIA

Internet:	bruce@cs.su.oz.au
Telephone:	+61-2-692-3272
Fax:		+61-2-692-3838

>From jeffs@soul.esd.sgi.com  Wed Mar 13 19:55:51 1991

It depends on your needs.  The JFS (AIX/6000's file system) does not do
fragmentation like the Berkeley FFS.  The blocksize is 4K.  Therefore a
one byte file takes 4K on disk.  With most FFS based systems the same
file will takek 512bytes or 1K.

If you have lots of small files (typical) than there are probably better
choices than the RS/6000.  For larger files I think the RS/6000 would
perform fairly well.

For general use though, and if cost is a priority, another system would
probably be better suited to your needs.

Good luck

jeffs@sgi.com
(ex-IBM employee)

>From tubsibr!schoenw  Thu Mar 14 09:49:31 1991
Organization: TU Braunschweig, Informatik, Bueltenweg, Germany
Cc: 
If you get any valuable answers I would like to get a summary. 
I would also like to hear about experiences using Unitree.

-- 
Juergen Schoenwaelder (...mcsun!unido!infbs!tubsibr!schoenw)
Technische Universitaet Braunschweig, Inst. f. Betriebssysteme & Rechnerverbund
Bueltenweg 74/75,  3300 Braunschweig, Germany, Tel. +0049 531 / 391-3246

--
Betty Jo Armstead              SVERDRUP Technology Inc.
21000 Brookpark Rd.Ms 142-2
Nasa Lewis Research Center
Cleveland Ohio 44135           From: xxbja@csduts1.lerc.nasa.gov

john@iastate.edu (Hascall John Paul) (03/16/91)

In article <1991Mar15.204401.28337@eagle.lerc.nasa.gov> xxbja@csduts1.lerc.nasa.gov (Betty Jo Armstead) writes:
}Here are the replys I received in response to the post on distributed
}servers.

   I missed the original request for this because the powers that be
decided that buying a used disk for News spool was a good place to save
money (har!).  Anyhow, from the other responses, it sounds like we are
running one of the larger fileserver setups.  We currently have 4 of the
following (probably 5 soon):

            DECstation 5000/200 CX grayscale
            16-24 MB memory
            an additional SCSI card
            8-10 x RZ57 (1GB each)
            TZK08 (really an EXABYTE)

with which we are serving 300+ DECstations of various types, ~1500 users
in a couple dozen buildings (heaven help us when we let the undergraduates
on!) with NFS, Mail, Print, and all the "Athena" services, etc...

   Our request for bids was answered by Apollo(HP), DEC, IBM, SGI & SUN
(I gather you can figure out who we selcted).

   We have had a goodly number of companies, universities, etc. through to
have a look and I am sure we could survive some more ;-)

John Hascall
--
John Hascall                        An ill-chosen word is the fool's messenger.
Project Vincent
Iowa State University Computation Center                       john@iastate.edu
Ames, IA  50011                                                  (515) 294-9551