Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!tut.cis.ohio-state.edu!ucbvax!hplabs!hp-ses!hpcuhb!hpindda!jack
From: jack@hpindda.HP.COM (Jack Repenning)
Newsgroups: comp.sys.hp
Subject: Re: a local swap disk for HP9000/300's ?
Message-ID: <4310058@hpindda.HP.COM>
Date: 13 Sep 89 18:31:00 GMT
References: <2295@umbc3.UMBC.EDU>
Organization: HP Information Networks, Cupertino, CA
Lines: 73

> I've seen several references saying that 6.5 can have a local swap
> disk, what do you think?

You can have local swap (although you can't have local filesystems).
This has been true since the very beginning of Diskless (6.0).

However, local swap is not necessarily a performance gain!  This is
because the difference between fast disks and slow disks is greater
than the difference between local swapping and remote swapping.  Since
the faster disks are generally bigger, it's more economical to share
them (by putting them on the server) - and not many single
workstations really need 304Mb of swap space, anyway.  

There are (at least) two ways you can gain performance by adding local
swap.  First, local swap on any client will probably help the
rootserver (even if the client gets slower).  Second, if there's so
much swapping going on, that swap requests are being queued up at the
server, then local swap might help.  A simple intuitive test of this
would be, is your cluster too slow when only one workstation is in
use, or only when two or three are used concurrently?

All of this assumes that you really are paging and swapping enough for
that to be your major performance concern, of course.  You might want
to use /usr/contrib/bin/monitor's "k" screen, to verify that your
workstations are really using significantly more swap space than their
(physical memory - kernel size).  You can also use the "g" screen to
see what demands you're putting on the disk bandwidth.

What can you reconfigure?  Well first, gut everything you don't need
from the kernels: client and server.  Second, play with the
configuration variables, to reduce tables as much as possible.  "As
much as possible," you'll have to discover for yourself - it depends
on what you do with the workstations.

Here are the tunings I've installed in my cluster.  They work for us,
but of course your needs may be entirely different.  (We do a lot of
X, mostly X11, mostly mail reading, editing, HP TeamWork - compiles
are run on S800 systems.)  The server tunings, in particular, are for
a 16M server.  If you have less memory, you'll want to pull back
"nbuf" considerably.  (But do tweak it up from the default: it's an
important tuning to improve filesystem access from the clients.)

* CONFIGURABLE PARAMETERS:
*
*	Tunings for X workstations
*	(Assumptions: one X user, maxo-piggo; maybe a few visitors
*	 from elsewhere, not using much the main user isn't)
maxuprc 64
maxusers 5
nproc	(20+MAXUPRC+((MAXUSERS-1)*8)+(NGCSP))
ntext	(20+MAXUPRC+(NGCSP))
*
*	Tunings for cluster server only
nbuf 512
ngcsp	(4*NUM_CNODES)
num_cnodes 10
server_node 1


-------------------------------------------------------------
Jack Repenning - Information Networks Division,
		 Hewlett Packard Corporation
uucp:   ... {allegra,decvax,ihnp4,ucbvax} !hplabs!hpda!jack
  or:   ... jack@hpda.hp.com
HPDesk: Jack REPENNING  /HP6600/UX
USMail: 43LN; 19420 Homestead Ave; Cupertino, CA  95014
Phone:  408/447-3380                     HPTelnet: 1-447-3380
-------------------------------------------------------------
Disclaimer: These opinions are not necessarily those of my employer.
In fact, my manager doesn't know I'm . . . Oh shoot!  Here he comes!
Quick!  Pop up that window with code in it!

Bye!