DATA CENTERS, NETWORK BOTTLENECKS,
AND THE MOVE TO PARALLELISM IN OPTICS
Optical communication systems that are ubiquitous in today’s Internet
and data transport networks have seen more than four orders of
magnitude increase in capacity since first deployments roughly thirty
years ago. Over this same period, their application has evolved from
being a high performance specialized technology reserved for the
Internet backbone to finding their way into personal computers and
home systems. Optical transceivers today are sold in volumes of 10’s
of millions per year for data center networks. This rapid expansion of
the use of optical systems and in particular the migration of optics to
the edge of the network is placing burgeoning requirements on the
scalability of optical systems.
From a performance standpoint, optical systems
are approaching the Shannon channel capacity
limit, which bounds the spectral density. As a
result, historical scaling methods through increases
in the bandwidth-distance product are coming to
an end. This trend, often referred to as the fiber
capacity crunch, is driving a shift to scaling through
parallelism. The reach and capacity improvements
of a single-amplifier-band fiber system are
reaching the point of diminishing returns due to
physical and practical limits. Further capacity gains
will need to come from multiple parallel systems
or integrated systems that incorporate multiple
fibers or amplifier bands. To keep systems
affordable, the main challenge in scaling
through parallelism is that the energy and
footprint need to scale in proportion
to the performance, i.e. exponential
traffic growth can only be supported
by exponential reductions in the
energy per bit and the area
per unit bandwidth. Focus
on these quantities is a
fundamental shift from
how systems have been
developed so far.