| |
| |
|
|
| |
PANEL
DISCUSSION | WEDNESDAY, AUG 25
What are the future
trends in high-performance interconnects
for parallel computers? |
| |
Panelists:
Jose' Moreira, IBM ; TJ Watson; Jarek Nieplocha,
Pacific Northwest National Laboratory; Mark Seager,
Lawrence Livermore National Laboratory; Craig Stunkel,
IBM; TJ Watson; Greg Thorson, SGI |
| |
|
| |
Ever-increasing
demand for computing capability is driving the construction
of ever-larger computer clusters, comprising compute
nodes integrated by a high-performance interconnect.
This network is in most cases the heart of the parallel
computer and defines its functionalities, influences
the design of the system software and determines
the actual performance of the machine.
The panelists will discuss the major design trends in high performance
networks.
|
| |
• |
Most
interconnection network provide some form of "intelligence" in
the
network interface. Do you expect this to become a central feature in the
future? Will it be possible to implement "network operating systems" in
the
network interface? |
| |
|
|
| |
• |
Do
you expect that optical networks will become widespread? |
| |
|
|
| |
• |
What
are the trends in latency and bandwidth? Networks
as Quadrics Elan4
and Cray XD1's already deliver 1.5 microseconds at MPI level. Is bandwidth
technologically free? Will the I/O interface be the bottleneck? |
| |
|
|
| |
• |
Will
native support for collective communication be a
central feature of a
high performance network? |
| |
|
|
| |
• |
BlueGene/L
has shown that a thermally-aware supercomputer can
be packaged
in a small space, with a chip that integrates processors and network
interface. Will future network interfaces be integrated with the processors
in the same chip? |
| |
|
|
| |
• |
How
can a high-performance network help to achieve fault-tolerance
in a
large-scale machine? |
| |
|
|