Size Does Matter!
(From the Age of Closed-Loop to the Age of Open-Loop)

Roch Guerin (guerin@ee.upenn.edu)
University of Pennsylvania

By many accounts, the Internet is out of control (SPAM, DDOS, DNS
attacks, etc.) and in spite of its touted simplicity and flexibility,
the sheer weight and complexity of existing protocol suites makes
introducing fixes to those problems increasingly difficult.  But by
the same token it keeps growing and allowing experimentation with and
deployment of new services.  The Internet is morphing with every
"killer app" that gets added, from the web, to Skype, to YouTube and
Joost.  There are glitches and setbacks, but the growth is there and
unabated in spite of repeated predictions of its pending doom, e.g.,
from a video over-dose, as many have recently warned against.  So on
one hand, the Internet is faced with a series of increasingly serious
and potentially crippling problems, and on the other hand it is
getting bigger and more diverse by the day with new services and new
types of users, e.g., wireless devices, being continuously added.
This clearly makes for a challenging situation where "fixing the
Internet" could be compared to trying to change wheels on an
accelerating freight train.  This is hard not only because solutions
need to aim at a moving target, but also because the fact that the
target is moving highlights that it is far from having reached a stage
where its deficiencies have sufficiently affected it and made it ripe
for either major changes or replacement.  As a matter of comparison,
one could argue that the phone network had been at a stand-still for
many years before the Internet came along and ultimately displaced it.
The current situation is clearly vastly different.

However, while the Internet's sheer size and its continuing growth are
the main contributors to the challenges one faces when attempting to
improve it, they also create opportunities in leveraging size as a
critical component to develop solutions addressing these problems.  In
particular, the magnitude of the Internet coverage and reach also
means that it offers an incredible diversity in terms of both
connectivity and the characteristics of that connectivity.  This
global diversity is further supplemented by the growing availability
of local diversity in the form of new access technologies, e.g.,
multi-homing, multi-band radios, etc.  In other words, in today's
Internet, there is increasingly a plethora of choices for realizing
connectivity between any two points.  As a result, as the Internet and
its diversity keep growing, the odds of the many connectivity options
it offers all going simultaneously bad are fast diminishing, and here
lies the key to developing successful solutions. Specifically, while
we should still strive to continue improving the stability,
correctness, security, performance, etc., of the many protocols and
technologies that play a role in delivering the ubiquitous
connectivity that has made the Internet so successful, devising
solutions that can keep up with rather than hinder its growth must
directly take advantage of that growth.

This has a number of important implications.  First and foremost, it
is critical to provide easy access to as well as leverage the
different connectivity options that may be available at any given
point in the Internet.  This affects routing protocols and in
particular a protocol such as BGP that is predicated on the selection
of a single (best) path, as well as the forwarding behavior of
routers.  For example, if diversity is viewed as desirable, it should
be incorporated in the path selection process to favor choices that
maintain or increase the number of forwarding alternatives available,
and conversely packets of an individual flow should preferably be
forwarded over as many distinct paths as possible rather than be all
mapped onto the same next hop as is commonly done in today's routers.

A second important implication is that both end-systems and routers
should incorporate mechanisms that seek to directly exploit diversity.
In particular, packet replication becomes an important functionality,
which when coupled with mechanisms that map different packet replicas
onto different forwarding decisions can deliver significant
improvements in performance and reliability, simply by taking
advantages of the diverse characteristics of the underlying paths
(they may not all be perfect, but they will most likely not all be
bad).  In a sense, this is no different from the age-old design
principle for building highly reliable systems from unreliable
components.  Specifically, the Internet is just like any other
complex, large-scale system, whose reliability can to some extent be
improved by making its components more reliable, but for which
redundancy alone can provide a solution that overcomes the intrinsic
challenge associated with scale.  

Last but not least, as is common when scale is a key factor, reactive
solutions are to be used only with caution.  This is not only because
the many possible root-causes behind observed changes can make
identifying the appropriate reaction difficult, but also because
large-scale reactions, e.g., from many users, can introduce problems
of their own such as instabilities and oscillations.  In other words,
open-loop approaches should and will play a major role in developing
solutions that can scale with the Internet.  This will be true for
both network control functions, e.g., the computation of backup paths,
and mechanisms at the level of end-users flows, e.g., by using
diversity coding as previously alluded to.

In summary, solving the Internet's problems can only be realized by
viewing its size as an advantage rather than a factor that renders
solutions more complex and difficult.  With size comes diversity, and
by exploiting diversity it is possible to develop solutions that
combine multiple weak links into one strong link.  More generally, it
is important to acknowledge that no matter what improvements one can
make to individual components, the reliability of any large-scale
system must involve some level of redundancy.  The focus should,
therefore, be on exploring how to best add and utilize redundancy
across the Internet and its users.