SUMMARY: Dataless Clients on a different subnet

From: Geoff Mellor (grm@star.le.ac.uk)
Date: Thu Sep 29 1994 - 12:35:55 CDT


I recently wrote:

>Dear Managers,
>
>We have a group of Solaris 2.3 machines - a server and some dataless
>clients. The campus network people have forced the situation upon
>us whereby one of the clients is now on a different subnet.
>
>To continue to run as before we need to set up a default gateway
>before attempting to mount /usr for single user mode.
>In the startup file /etc/rcS.d/S30rootusr.sh, there is
>provision to do this, if we get a static copy of route and place it
>in /sbin.
>
>So I contacted my local Sun support who said that they don't have
>a static version of route because this method of client serving
>is unsupported. They reckon that performance is very poor and
>didn't want to get involved. They said that the problem might get fixed
>in a later Solaris release.
>
>So, can anyone tell me:
>
>1) Is performance really that poor over subnets ?
>2) If so, why ?
>3) Can anyone point me to a static version of route so I can try it
> myself ?

>I will summarise.

>Geoff Mellor Internet: grm@star.le.ac.uk
>Systems Administrator Phone : +44 (0)116 252 3599
>Dept of Physics and Astronomy Fax : +44 (0)116 252 3311
>University of Leicester

Thanks to the following who replied:

D.G.Jones@scuna.dircon.co.uk
wolfgang@sunspot.nosc.mil
stlee@alc.com
glenn@uniq.com.au
Eckhard.Rueggeberg@ts.go.dlr.de
paulo@s2.dcc.unicamp.br

Basically, it was generally felt that serving across routers
is a bad idea and that I should avoid this. My current client
does not have enough resources to be a standalone machine
but in a couple of months it will be possible to obtain enough
local diskspace to do this.

In the meantime, a solution had to be found. I edited the netmasks
file and opened up the subnet mask to 255.255.0.0 for the duration
of the boot. It was then explicitly closed down again to 255.255.255.0
in /etc/rc2.d/S72inetsvc. This avoided the need to obtain a statically
linked version of route.

Performance is not great but it will get us by until we can create
a standalone machine in the other subnet.

I include the replies I received below:

From: Derek Jones <D.G.Jones@scuna.dircon.co.uk>

You may end up overloading your router. Serving across routers is a
bad idea generally speaking if there is any modicum of traffic.
What is the context switching rate of the router? How much other
traffic does it see? What are the relative speeds
of the router and the dataless client? etc. etc.

My fix would be to unforce the situation 8-) (But having until recently
worked as sysadmin in a Uni myself, I quite understand why it doesn't
always work . . .) Sorry that's not much practical help.

Kind regards

Derek Jones

From: "Lewis E. Wolfgang" <wolfgang@sunspot.nosc.mil>

Bandwidth limitations are the main reasons inter-subnet
client support doesn't work well. We've tried it over
a bridge and we were plagued with many timeout errors,
balky performance, etc. And security considerations figure
in when you want to boot diskless clients.

Better to set the orphaned machine up in standalone mode.

Good luck,
Lewie

From: stlee@alc.com

The reason for subnetting is to reduce network traffic. A dataless client
generates a large amount of network traffic. By running a dataless client
over the router you are defeating the reason you were subnetted in the first
place. That may not bother you but in the long run doing things that way
will force the sysadmins to further subnet you before they should have to,
not a fun thing for anybody.

You might consider getting a second card for this machines nearest neighbor and
setting that machine up as a gateway.

> So, can anyone tell me:
>
> 1) Is performance really that poor over subnets ?

Depends on the hardware used to subnet.

> 2) If so, why ?

Any packet that needs to cross over has to be checked by the hardware, it then
looks at each packet to determine if it should go.

From: Glenn Satchell - Uniq Professional Services <glenn@uniq.com.au>

You can get a copy of route.c from the XKernel package, check archie
for a site near you. This is the most common source, although I'm not
sure if it is ported to Solaris 2 yet.

As for performance, the goal of subnetting is to keep traffic local to
one subnet and only relatively small amounts of inter-subnet traffic
gets routed. The performance will depend on your router and how fast
it can route packets. There will be some degradation, but how much can
only be found out by trying it.

From: Eckhard Rueggeberg <Eckhard.Rueggeberg@ts.go.dlr.de>

> 1) Is performance really that poor over subnets ?

No it won't be poor at all if you use cachefs. Without cachefs, it
depends on the speed of the router(s).

> 3) Can anyone point me to a static version of route so I can try it
> myself ?

I never tried it myself, but if you want to cachefs /usr, you have to
have a local mini-/usr which contains the cachefs tools. And if it
contains the route binary, too, this should work fine for you.
Sorry for being inprecise, but I don't remember what you have to have
on a local /usr (no need for a partition, simply a directory in the
/ partition which gets overmounted).

>From paulo@s2.dcc.unicamp.br Thu Sep 29 11:04:29 1994
Date: Mon, 26 Sep 1994 15:23:19 +0300
From: Paulo Licio de Geus <paulo@s2.dcc.unicamp.br>
To: Geoff Mellor <grm@star.le.ac.uk>
Subject: Dataless Clients on a different subnet

I've done the some thing for other purposes. A dataless client needed
a static route to be able to deal with a second ethernet interface,
but that was with SunOS 4.1.3.

Instead of a static route, I've added to the "/" filesystem a /usr
directory containing lib/libc* and a few other libraries, so that the
support for shared libraries and the specific library needed by route
are available immediately. That's enough to get the machine booting
in my case. That machine no longer works under this config, so I
can't reproduce the details right now. The /usr dir actually gets
covered by the full /usr when it's mounted via nfs.



This archive was generated by hypermail 2.1.2 : Fri Sep 28 2001 - 23:09:10 CDT