M.I.T. Laboratory for Computer Science IEN 143
March 11, 1980
Environment Considerations for Campus-Wide Networks
by Jerome H. Saltzer
"The Campus Environment" is a name proposed here to identify a
particular set of physical properties, geographical extents, data
communication requirements, administrative relationships, and needs for
flexibility that characterize our university campus. With only minor
exceptions they equally apply to a corporate site, a government complex,
or another university. This note discusses seven characteristic
properties of this campus environment. These seven properties provide a
basis for design decisions for a data communication network to span a
campus. As will be seen, the properties of this environment are quite
different from those of a single building, or of a nation-wide,
common-carrier-based network.
Seven Properties of the Campus Environment
1) It has a geographical extent beyond a single building, but within a
single political and administrative boundary that permits
transmission media to be installed without resort to a common
carrier.
This first property is essential, so as to allow exploitation of
low-cost, high-bandwidth communication technology. With current
technololgy and prices the difference in costs between communicating
over privately installed equipment and using common carrier facilities
can be a factor between 10 and 100.
2) Within this geographical area, a large number of nodes–that is,
2
computers, data sources, and data sinks--require interconnection.
Today the number of such nodes may be in the range of ten to one
hundred. Looking ahead to the advent of desktop computers, one may
be faced with from a few hundred to several thousand nodes by the
end of the next decade.
The combination of the previous two properties seems to make it
inevitable that local interconnect technologies such as the ETHERNET,
CHAOSNET, L.C.S. Ring net, HYPERCHANNEL, or MITRENET cannot by
themselves completely accomplish the required interconnection, since all
such technologies that have so far been demonstrated have limitations on
distance on the order of a thousand meters and limitations on node count
on the order of a hundred nodes. Thus one would expect to use those
technologies to attach clusters of nodes into subnetworks, for example
all the nodes in a single building, and then install interconnections
(gateways) among these subnetworks. For our own campus, one might
envision by 1990 as many as 100 subnetworks each comprising an average
of, say, 100 nodes. Subnetworks and gateways introduce the problem of
how to route a message from a source node through a series of
subnetworks and gateways, so that it ends up at a desired target node.
3) Administratively, there exist forces both for commonality and for
diversity of network attachment strategies. The primary force for
commonality is a desire to be able easily to set up communications
between any pair of nodes on the campus. The primary force for
diversity is that the choice of a computer, data source, or data
sink typically pre-determines the technology of the network to
which it must be attached, because off-the-shelf network hardware
3
for that node may be available in only one technology. Further,
some applications may have special requirements for some
connections (e.g., high bandwidth) that can be met only with a
particular network supplier's equipment, yet still need occasional
"ordinary" connections to nodes elsewhere. Thus the emerging
diversity of local networks will continue, and probably increase,
rather than decrease, with time.
4) The worldwide academic, commercial, and regulatory community has
not yet reached anything resembling a consensus on how functions
should be divided. Arguments range over issues ranging from obscure
matters of taste, through fundamental technical disagreements about
which requirements should have priority in design, to alternative
opinions of the directions that communication technology is moving.
Many different and competing standards have been proposed, and one
can find in the literature a good technical case against any one of
them. One must anticipate that these arguments will be reflected
internally in the campus environment, in the form of a diversity of
protocols and standards, and particularly in the requirement that
any mutually consenting set of nodes be able to carry on
communication with one another using a protocol that no one else
has ever heard of, much less agreed to. [Imagery borrowed from a
Chaosnet working paper by David Moon.]
This fourth requirement suggests strongly that any network
interconnection strategy that must be implemented today should have a
campus-wide lowest layer of protocol that accomplishes datagram passing
between any two nodes while making an absolute minimum number of
4
assumptions about the nature of the higher-level communications that are
taking place or the policy of network administration. Some typical
assumptions that should be avoided unless an unusual opportunity is
obvious are: what level of reliability/delay tradeoff is appropriate;
how routing should be optimized; fragmentation/reassembly strategy; flow
control requirements; addressing plan; and particular network topology.
5) Because a data communication network is a campus-wide service,
there will be no single user or user group with a wide-enough
interest to administer the entire network. This means that network
administration will either be done by a haphazard confederation of
special interest groups or else by a chronically underfunded
central service organization modeled on the one whose role is to
minimize telephone costs.
In either case, this property places a requirement on the network
interconnection technology that it be robust and self-surviving to every
extent imaginable. Trouble isolation must be easy to accomplish and easy
for individual users to participate in if they are so inclined, because
trouble isolation and repair may involve multiple administrations.
Simplicity of operation of gateways is important, so that operation can
be completely unattended for long stretches of time. A network design
approach that requires close monitoring is undesirable.
6) The topology of subnetwork interconnection will be administered
partly with central planning and partly without. This property
arises from two needs: First, a "dependable" set of gateways that
one can expect to exhibit predictable and stable properties is an
essential backbone to a useful service. A centrally planned and
5
administered set of gateways would provide this dependability.
Second, whenever a node finds that for some reason it is attached
to two subnetworks, it may find that it is useful in some of its
applications to serve also as a gateway between the subnetworks;
yet it may not want to take on the official responsibility of being
a publicly available gateway. Another example of a gateway that is
not centrally administered may arise if some particular application
needs, and has purchased the gateway equipment to support, a path
through the network with special properties of delay, reliability,
bandwidth, or privacy. The person or organization that has
purchased the special gateway equipment may not be prepared or
willing to allow public use of it. Alternatively, a user may wish
to avoid use of a sometimes troublesome gateway that is claimed by
its owner to be perfectly operating.
7) External networks such as TELENET, the ARPANET, TYMNET< XTEN, SBS,
or A.C.S., may be attached to some nodes, and some of those nodes
will serve as gateways between the campus network and the external
networks. In some cases, the external network will be used simply
as a "long link" in the campus net. In other cases, facilities
within the campus net will set up communication paths to services
having no other connection with or knowledge of the campus net.
Both kinds of cases require careful consideration of the
interactions between internal and external network properties.
Note that the campus environment has all these properties only if we
assume the technological opportunity mentioned in point one: that
low-cost hardware and media can provide communication paths in the range
6
from 1 to 10 Mbits/sec. between any two points within the campus.
Availability of interconnect media and subnetworks with this bandwidth
has been demonstrated in several forms. Gateways that operate with such
bandwidths may be harder to construct, and that concern is one of the
considerations involved in developing a campus-wide net. Individual
nodes that can sustain these data rates for very long are likely to be
rare; software often limits the rate at which a mode can act as either a
data source or data sink. Instead, the high bandwidth technology is to
be exploited in two ways:
1) to provide enough capacity to handle the aggregate demand of
many lower-bandwidth sources and sinks of data.
2) non-optimal strategies that are relatively simple to implement
or administer can be considered; it is not a requirement that
every bit of the available band- width be optimally utilized.
The availability of high bandwidth, together with lack of a requirement
to use that bandwidth efficiently, is probably the most fundamental
technical difference between the "campus-wide network" and the
commercial long-haul data communication network, a difference that can
lead to significantly different design decisions. Future notes in this
series will explore some of these specific technical design
consequences.