Next: The Future of
Up: SuperComputing-94 --- Networking Technologies
Previous: Ethernet
This section of this report on networking technologies at
SuperComputing '94 highlights some of the comments made at the
Gigabit Networking: State of the Art and Applications Roundtable.
Many of the comments reported here appear as sound-bites --- single,
notable catch-phrases. Given the short time for presentations by the
members of the roundtable, the talks often made only one or two points
and were presented in the same manner as sound-bites. Those
presenting at this roundtable where highly technical, knowledgeable
people in the field of high-performance networking to support
supercomputing. All comments reported here are from the panelists
unless otherwise specified.
- Serial HiPPI has enabled this high speed networking technology
to move out of the machine room, and move onto the desktop.
- ATM will be the WAN technology in the near future, and ATM
will drive other technologies out of the WAN market within the next
five years. ATM will continue to increase its market share in the
LAN market as global networking access increases and users want faster
access to resources located throughout the world.
- The next generation of research must examine ways to reduce
both hardware and software latency in networking products. At present,
latency in distributed computations is the most significant issue ---
ATM is presently only 50% faster than Ethernet --- entirely due to
hardware and software latency within the workstation.
- Sandia National Laboratory is concentrating its networking
investment on ATM technology. Sandia has strong interest in
heterogeneous computing over the wide area. Sandia has found that
software latency is one of their biggest problems, and they are
examining ways to optimize the entire software stack to minimize
latency and improve performance. Sandia is looking for ATM switches
with high capacity, gigabits-per-second, to handle their network
traffic. Sandia claims that it has not been difficult to obtain the
massive bandwidth to interconnect their enterprise networks, but the
cost has been high. At present Sandia is paying $500 per month per
mile for connectivity from commercial sources. The only thing that
could be worse than those high costs would be to have to develop and
maintain those communications capabilities in-house.
- The single killer application of the future is the lawyer
application. Legal concerns may minimize data compression for
medical images, or anything other than end-user entertainment. A
legal case already has been filed because image compression destroyed
the phenomenon of interest on a medical image.
- A representative of the Ford Motor Company discussed their
corporate enterprise network(s). Ford presently has world wide
TCP/IP, SNA, ... , etc. networks with the interest in consolidating to
one networking technology. Ford has a supercomputing center in
Dearborn, MI equipped with (among other computers) massively parallel
computers from MasPar. Ford has made a commitment to build world
cars and has instituted a corporate policy to design automobiles over
networks sharing data and using available remote resources. Designers
in Australia have access to the most available computing power, due to
their work times in the global day. (Authors note --- remember that
Ford is a private company, competing in a world economy, using
state-of-the-art networking and distributed computing. Ford's
networking plans appear as ambitious as Sandia's.)
- HiPPI is RS-232 (a serial technology) on steroids. We can get
gigabits-per-second to a workstation today. Eventually, we will be
able to use the workstation to do something with that data.
Presently, the entire workstation is required to support a
gigabit-per-second interface. Even more capable workstations are
required if we are to use such large amounts information in
constructive calculations or transaction processing.
Next: The Future of
Up: SuperComputing-94 --- Networking Technologies
Previous: Ethernet
David P. Koester
Sun Oct 22 13:05:27 EDT 1995