High-Performance Interconnection Networks for Cluster Computing
Wednesday, August 20, 6:30-8:00p
Stanford University, Memorial Auditorium Panel Organizer and Moderator
Panel Questions Contributed by Panel Members - Greg Pfister (IBM Austin)
- Kevin Daierling (Mellanox)
- Moray McLaren (Quadrics)
- Wu-chun Feng (LANL)
- Ron Brightwell (Sandia)
The panel will address two major sets of questions:
- Which interconnect is the best for high-performance computing and why?
- What are the future trends in high-performance networking, and what are the implications of these trends?
Specific questions include:- Are quantitative measures of latency and bandwidth enough tocharacterize a network interconnect, what other ways should we beevaluating interconnects?
- Will the "status quo" in networking continue? That is, Ethernet as acommodity interconnect that is also used as a cheap commodity solutionfor clusters with InfiniBand, Quadrics, and Myrinet "relegated" to high-end and more costly clusters.
- What assumptions must interconnects make about the underlyingarchitecture (or what assumptions would they like to make)? PCI-X? PCIExpress, Hyper Transport?
- In five years, how will today's interconnects evolve and/or competein high-performance computing?
- InfiniBand started out as a high-performance I/O technology but hasevolved into a general network interconnect for high-performance clusters. Will it replace Myrinet or Quadrics as the costlier high-performance interconnect for high-end clusters?
- What features and improvements are needed in communicationsubsystems to build next generation clusters? (network hardware,communication layers, libraries, programming models, etc.).