|
As technology pushes the Petascale barrier, and the next generation Exascale
Computing requirements are being dened, it is clear that this type of computational
capacity can be achieved only by accompanying advancement in
communication technology. Next generation computing platforms will consist
of hundreds of thousands of interconnected computational resources. Therefore,
careful consideration must be placed on the choice of the communication
fabrics connecting everything, from cores and accelerators on individual platform
nodes, to the internode fabric for coupling nodes in large-scale data
centers and computational complexes, to wide area communication channels
connecting geographically disparate compute resources to remote data sources
and application end-users.
The traditional scientic HPC community has long pushed the envelope
on the capabilities delivered by existing communication fabrics. The computing
complex at Oak Ridge National Laboratory has an impressive 332 Gigabytes/
sec I/O bandwidth and a 786 Terabyte/sec global interconnection
bandwidth [305], used by extreme science applications from \subatomic to
galactic scales" domains [305], and supporting innovation in renewable energy
sources, climate modeling, and medicine. Riding on a curve of a thousandfold
increase in computational capabilities over the last ve years alone, it is expected
that the growth in both compute resources, as well as the underlying
communication infrastructure and its capabilities, will continue to climb at
mind-blowing rates.
Other classes of HPC applications have equally impressive communication
needs beyond the main computing complex, as they require data from remote
sources | observatories, databases, remote instruments such as particle accelerators,
or large scale data-intensive collaborations across many globally
distributed sites. The computational grids necessary for these applications
must by enabled by communication technology capable of moving terabytes of
data in a timely manner, while also supporting the interactive nature of the
remote collaborations. |