we offer the Deployment Scalability
Trade-off (DST): There is an inherent
correlation between deployment scalability of a system with a given spanning layer and the weakness, simplicity, generality, and resource limitation
of that layer’s specification.
The original motivation for creating
a formal model of layered systems was
to better understand end-to-end arguments in the context of the hourglass
model. We sought to explain and guide
efforts to generalize shared network resources while addressing the intuitive
design principle that the requirements
of scalability had to be the primary
and overriding constraint. The DST is
a candidate as a more general design
principle that situates end-to-end arguments in a complex space of design
criteria. The Hourglass Theorem is a
first step in an explanation of the role
of logical weakness in the DST.
Examples and Applications
Giving an account of an application of
the Hourglass Theorem can be tedious.
The antecedents of the theorem require the definition of the specification
language, a program logic, all acceptable programs and the set of necessary applications N. This presentation
will be restricted to giving a less formal
account of the implications of the DST.
Tree building in IP multicast. Global Internet routing is made possible
through the use of interoperable approaches to internal and external routing within and between local networks.
The metrics assigned to individual links
by network administrators are somewhat arbitrary, but when used as inputs
to a combination of shortest path algorithms (interior gateway protocols) and
a policy driven peering protocol based
on commercial agreements (the Border
Gateway Protocol), the result is often
acceptably similar to some intuitive notion of efficiency.
Multicast routing is much more
complex. An IP multicast group is
based on tree-structured forwarding,
The design principle sought to
examine is that thinness of the span-
ning layer is correlated with greater
success in its adoption and longevity.
This is sometimes expressed as the
system exhibiting scalability. How-
ever, scalability means little if we do
not specify the attribute in which we
desire the system to exhibit an ability
Deployment scalability is used to
characterize a spanning layer being
adapted to finding success in the form
of widespread adoption. Deployment
scalability is intended to imply the
kind of “viral” adoption that the Internet and Unix spanning layers have exhibited. This definition is proposed as
a (admittedly imprecise) characterization of success in global infrastructure
service interface design.
Definition 9. Deployment scalability
is defined as widespread acceptance,
implementation, and use of a service
The notion of deployment scalability is introduced in order to have a vocabulary for expressing the goal that
is implicit in the design of a spanning
layer for global infrastructure.
For example, in describing the role of
the Internet’s thin waist, Peterson and
Davie8 state “The hourglass’s narrow
waist represents a minimal and carefully chosen set of global capabilities that
allows both higher-level applications
and lower-level communication technologies to coexist, share capabilities
and evolve rapidly.” It is the meaning of
“minimal” and “carefully chosen” that
we are trying to characterize.
In terms of the formal model of layered systems, we suggest that having
many possible supports and many possible implementations is correlated with
the goal of deployment scalability. The
Hourglass Theorem would then extend
this to a correlation between minimal
sufficiency and deployment scalability. We have given some informal arguments to support similar relationships
between other aspects of a thin spanning layer and deployment scalability.
Each of these aspects can be evaluated in isolation, but in the service of
our original motivation to link the thin
waist to a general notion of scalability,
shown a structural
of the spanning
in the set of