Path: utzoo!attcan!utgpu!jarvis.csri.toronto.edu!cs.utexas.edu!usc!jarthur!polyslo!rnicovic
From: rnicovic@polyslo.CalPoly.EDU (Ralph Nicovich)
Newsgroups: comp.protocols.tcp-ip
Subject: Re: subnetting on non-byte boundaries
Message-ID: <1989Dec13.201716.26369@polyslo.CalPoly.EDU>
Date: 13 Dec 89 20:17:16 GMT
References: <50277@srcsip.UUCP>
Reply-To: rnicovic@polyslo.CalPoly.EDU (Ralph Nicovich)
Organization: Cal Poly State University -- San Luis Obispo
Lines: 69

In article <50277@srcsip.UUCP> jkimball@src.honeywell.com () writes:
>
>Time to call on the Wisdom of the Net . . .
>
>We have a Class B network.  We've been following the strategy which all the
>examples in the manuals depict: subnetting at the byte boundary, using the
>third octet for our (internal) network ids and the fourth octet for the
>host id.  
>
>We've also been keeping our backbone as one logical network (actually three
>ethernet segments joined by learning bridges).  On that backbone network,
>we are approaching 250 hosts.  And we expect to need another 75 or so
>additional IP addresses within a few months. Ooops.
>
>Looks like our options are:
>
>  Option 1:  Modify our subnetting scheme.  Use 6 bits for internal
>     network id, and use 10 bits for host id.  We've kept our network
>     id's in the high bits of the third octet, so the low bits are free and
>     can be reassigned to the 'host part'.
>
>     Worry 1.1:  The only examples in the manuals show subnetting on byte
>       boundaries.  Will a 6/10 (vs 8/8) bit-split really work?  (Context:
>       lots of Suns running 4.0.3;  VMS VAXen running CMU-TEK 6.4; some
>       random Apollos and HPs that we don't care about, much.)
>
Here we have split ours with 4/12 since we have lots of terminal servers and 
wanted 4k addresses on one network. One of these logical subnets can then
go thru a router and change the mask on the otherside to 8/8. This way
we can hace lots of small subnets and a few big ones. We have had no problem
with devices such as the SUN's where you can set the mask at the bit level.
However there are some machines (HP STARLAN on A HP-3000) for instance that
won't support this since they simply ask what class you are when you install.
In it's case you must put it on the other side of a router that sets the
mask to class C.
>     Worry 1.2: Can we really expect to keep increasing the number of hosts
>       on our backbone network at this rate?  When will performance
>       problems set in?
>
When is very dependent on what you are dooing. If you don't use routers
broadcast packets will start to be a problem. Every device must look at
every broadcast packet and every packet addressed to itself. Even though
broadcast may be 3 percent of overall trafic it may be 80 % of what any
individual device must look at. Soon devices can't keep up with trafic
retransmition occurs and trafic goes up somemore....
>  Option 2:  Play games with routing.  Put two networks on the same
>       backbone, both with metric 0.
>
>     Worry 2.1: Is this some sort of unsupported kludge -- will it work,
>       for our mix of hosts?
I have tried this and it does work, it does not solve any trafic problems
however. With some devices that don't have all the nice berkley things
like metrics you still need a router with two ethernet cards connected
to the same ethernet. Trafic now doubles for some devices since it must
send packets thru this router. It is lible to give you heart burn.
>
>     Worry 2.2: If we should in the future attach some gateways to that
>       backbone cable, would they die of terminal confusion?
>
Probably most good gateways won't die (I used a CISCO Enet-X.25 it was ok)
But trafic will become an isue.

>  Option 3:  Bite the bullet and buy some hardware.  Replace the learning
>     bridges with true gateways, making the three segments be three networks.
>
>
>Has anyone done Option 1 or Option 2?  What would you recommend?
>
Recomendation: BITE THE BULLET.