[comp.protocols.nfs] One file server or several?

Chris.Rusbridge@levels.sait.edu.au (Chris Rusbridge) (11/07/89)

Apologies if this is not the right group for this query.

We are planning on buying a middle-sized UNIX system as a campus
server (aiming for 64 interactive users running edit, compile, run
and/or relational database work, and 16 diskless workstations, with 2
GByte disks or thereabouts). It has been suggested that we would be
better off buying two or more smaller systems instead. I guess there
seems on the face of it to be a lot more CP power in several
SPARCstation 1s than in one SPARCstation 370, for example.

But one consideration here is that we do not want our users to have to
be aware of where their files are (in the sense of having to remember
that their Pascal prac is on sysa but their 4GL prac is on sysc). The
users are also very mobile, and will be logging in from different
terminals at different times (probably mainly via telnet).

In the VMS world I am more familiar with, this can be achieved with a
VAXcluster. Can a similar result be achieved with UNIX (presumably nfs
and (whisper) yellow pages)? And what are the implications for
performance, Ethernet load and security?

Chris Rusbridge

Academic Computing Service Manager, SA Institute of Technology
ACSnet:         Chris.Rusbridge@levels.sait.oz [.au]
InfoPSI:        Chris.Rusbridge@sait.edu.au     (DTE 505282622004)
Phone:          +61 8 343 3098  Fax: +61 8 349 6939
Post:           Th

iarocci@Apple.COM (John Iarocci) (11/14/89)

In article <2669@levels.sait.edu.au> Chris.Rusbridge@levels.sait.edu.au (Chris Rusbridge) writes:
>Apologies if this is not the right group for this query.
>server (aiming for 64 interactive users running edit, compile, run
>and/or relational database work, and 16 diskless workstations, with 2
>GByte disks or thereabouts). It has been suggested that we would be
>better off buying two or more smaller systems instead. I guess there
>seems on the face of it to be a lot more CP power in several
>SPARCstation 1s than in one SPARCstation 370, for example.
>

I would go along with the suggestion to split up the load over
two or more sparcstations set up as servers than to depend on
one machine. By using NFS you won't have any problems with regards
to users having to remember where there files are located. Besides,
the 16 diskless clients will perform much better if they are not all
booting/swaping off of one machine.

Good luck.

					John Iarocci

iarocci@apple.com	apple!iarocci		(408) 974-3657
-- 
iarocci@apple.com	apple!iarocci		(408) 974-3657

smaug@eng.umd.edu (Kurt Lidl) (11/14/89)

In article <36451@apple.Apple.COM> iarocci@Apple.COM (John Iarocci) writes:
>In article <2669@levels.sait.edu.au> Chris.Rusbridge@levels.sait.edu.au (Chris Rusbridge) writes:
>>[user level tasks deleted]
>>It has been suggested that we would be
>>better off buying two or more smaller systems instead. I guess there
>>seems on the face of it to be a lot more CP power in several
>>SPARCstation 1s than in one SPARCstation 370, for example.

>I would go along with the suggestion to split up the load over
>two or more sparcstations set up as servers than to depend on
>one machine. By using NFS you won't have any problems with regards
>to users having to remember where there files are located. Besides,
>the 16 diskless clients will perform much better if they are not all
>booting/swaping off of one machine.

Another thing to consider is the actual ethernet chip in the machine
you are buying.  The SparcStations have the LANCE ethernet chip
(made by AMD) and the SparcServers have the Intel ethernet chip -- you
can guess who makes this one.  The performance of the two chips
is NOT even close to the same.  The le0 interface on the suns can
really make mince-meat of an ie0 interface on the fileserver.  When
this happens, NFS will do nasty things like swallow all your CPU
time.  Gory details available by E-mail.

>Good luck.
You'll need it.

>iarocci@apple.com	apple!iarocci		(408) 974-3657

--
/* Kurt J. Lidl (smaug@eng.umd.edu) | X Windows: Power Tools */
/* UUCP: uunet!eng.umd.edu!smaug    | for Power Fools        */

picard@ilog.UUCP (Yves Picard) (11/16/89)

In article <2669@levels.sait.edu.au> Chris.Rusbridge@levels.sait.edu.au (Chris Rusbridge) writes:
>Apologies if this is not the right group for this query.
> [...]
>But one consideration here is that we do not want our users to have to
>be aware of where their files are (in the sense of having to remember
>that their Pascal prac is on sysa but their 4GL prac is on sysc). The
>users are also very mobile, and will be logging in from different
>terminals at different times (probably mainly via telnet).
you might achieve the first goal by use of symbolics link:
suppose you have a dir called /udd (how original)
you make the following empty dir 
/udd/foo
/udd/bar
/udd/nil
...
where /udd/foo and /udd/bar linked to server1:/foo , server1:/bar
and /udd/nil linked to anotherserver:/nil
(as a matter of fact thing are no so easy but anyway ...)
the second goal (login from everywhere) is more or less difficult:
we are using a network of 40 machines with 4 main servers and 13 kinds
of computers (sun3 sun4 sun386 apollo hp9000 ibmrt ...) , we 'just'
copy every night "critical" files from a main server trough the net and
approximatively 38 out of 40 machines are working quiet well with it.

however if you net is homogenous you might try YellowPages.

	Jean Yves (john eaves )

PS: i know the stuff about /udd above isn't exact. more complete info
on required. no charge. no refund. no dogs allowed.
========================================================================
E-mail picard@ilog.ilog.fr

ksp@swrinde.nde.swri.edu (Keith S. Pickens) (11/18/89)

In article <1989Nov14.150708.2813@eng.umd.edu> smaug@eng.umd.edu (Kurt Lidl) writes:
>Another thing to consider is the actual ethernet chip in the machine
>you are buying.  The SparcStations have the LANCE ethernet chip
>(made by AMD) and the SparcServers have the Intel ethernet chip -- you
>can guess who makes this one.  The performance of the two chips
>is NOT even close to the same.  The le0 interface on the suns can
>really make mince-meat of an ie0 interface on the fileserver.  When
>this happens, NFS will do nasty things like swallow all your CPU
>time.  Gory details available by E-mail.

The 4/370 (SparcStation/SparcServer 370) uses the LANCE ethernet chip.
This same cpu is used in the SparcServer 390.  Older Sun-4/2X0 systems
used the the Intel ethernet chip.  The newer machines have been "improved".

	-keith
	 ksp@maxwell.nde.swri.edu