SUMMARY: client/server performance

From: Phillip Thomas (pthomas@netcom.com)
Date: Sat Mar 13 1993 - 01:29:35 CST


I posted these questions last week:

> I maintain a small network of Suns used for ECAD and MCAD applications.
> At present we have a Sun 4/470 used as an NFS file server and 8 Sparc2's as
> client workstations. We want to upgrade the clients to Sparc10's and add
> more clients as well. I have a couple of questions about performance:

> 1) Does an NFS file server need to have as fast a CPU as it's clients?
> If we upgrade the clients to Sparc10's do we have to upgrade the
       4/470 too?

> 2) I realize this is probably application dependent but is there a method
> to estimate how many clients a particular NFS server can handle?

> I would appreciate any input on this. In particular, real life examples of
> other peoples configurations would be helpful.

The majority of people replied that there is no problem having an NFS server
with a less capable CPU then it's clients. Many people have exactly that.
People agreed that the important characteristics for an NFS server are fast
I/O and network performance (several people sited the architecture of Auspex
NFS servers as an example).

Suggestions to enhance NFS server performance were:

1) Add memory to the server to increase disk caching and enhance NFS reads.
2) Use an I/O bus and disk drives with fast I/O capability (fast SCSI,
   synchronous SCSI, quick drive access time, etc.)
3) Use NFS enhancements such as Prestoserve, eNFS, NC400
4) Distribute the load on disks more equally if needed.
5) Use disk striping
6) Read the book "Managing NFS and NIS" by Hal Stern

Most people agreed that estimating client/server ratio is a tricky thing
and depends entirely on the applications run on the clients. The most common
suggestion was to monitor the NFS traffic to the server, using tools such as
nfsstat and nfswatch, and try to characterize the applications which generate
the NFS traffic. Many people just add clients until the server slows down and
then try to enhance the server (or get another server).

The following are actual configurations:

---------------------------------------------
From: Patrick O'Callaghan <poc@usb.ve>
---------------------------------------------
...we run a teaching/research lab with 16 clients going
to one server (an SS2). Writes are around 3% of the NFS ops we get.

---------------------------------------------
From: phil@dgbt.doc.ca (Phil Blanchfield)
---------------------------------------------
Here we have 10 dataless SS2 systems on a 670 and it is NEVER busy.
We even have one guy run Openwindows on it everyday.

---------------------------------------------
From: glenns@sa-cgy.valmet.com
---------------------------------------------
Just a brief note - our local system is based on a small group
of machines serving as NFS servers. We have a pair of DEC 2100
units and a couple of Sun SPARC (IPXs) doing this job. It seems
to work not too badly considering that there are in excess of
100 client machines at any given time. (But then again, our
actual NFS load is probably relatively light per host).

---------------------------------------------
From: aad@lovecraft (Anthony A. Datri)
---------------------------------------------
In a former life I had ~250 Sun 3/[56]'s, ~40 Convex machines, and ~20 Sun 4's
mounting a single filesystem from a 12-meg 3/160. Response time wasn't
amazing,but it served its purpose well.

---------------------------------------------
From: xcpgzm@atom.oryx.com (Phil Meyer)
---------------------------------------------

We have a SUN 690mp with two cypres CPUs serving 100 diskless (with local swap)
clients. We will run up the number of clients to 150 by the end of the month.
We use fast routers (alantec) in front of the 690, with FDDI between the routersand the server. This is just about the 'edge' of what the 690 can do. If
the powers that be will let us, we will either upgrade the 690 with two model
52 CPUs, and Solaris 2.2, or maybe even a SparcCenter 2000. But for now, our
two year old 690mp is handling the load better than I would have thought.

Thanks to:

clep@asdi.saic.com (Al Clepper )
dsnmkey@guinness.ericsson.se (Martin Kelly)
eckhard@ts.go.dlr.de (Eckhard Rueggeberg)
julian@syd.dwt.csiro.au (Julian Dryden)
<poc@usb.ve> Patrick O'Callaghan
<cal@soac.bellcore.com> Christian Lawrence
walfish@jpmorgan.com (Joshua Walfish)
ajs6143@eerpf001.ca.boeing.com (Andy J. Stefancik)
phil@dgbt.doc.ca (Phil Blanchfield)
<glenns@sa-cgy.valmet.com> Glenn Shaw
era@ncar.ucar.edu (Ed Arnold)
ups!kalli!kevin@fourx.Aus.Sun.COM (Kevin Sheehan {Consulting Poster Child})
shj@ultra.com (Steve Jay {Ultra Unix SW Mgr})
xcpgzm@atom.oryx.com (Phil Meyer)
mxsq@hercules.sa.mdss.pge.com (Michael Scott - SVR Sys. Admin)
bb@highland.com (Robert Bruce)



This archive was generated by hypermail 2.1.2 : Fri Sep 28 2001 - 23:07:36 CDT