9/20/2006 11:00am-12:00pm ECEE 1B28
|
A Simple, General Throughput Model for Interconnection Networks
National Science Foundation and University of Southern California
Interconnection networks are designed to transfer the maximum amount of
information within the least amount of time so as not to bottleneck overall
system performance. The design and evaluation of interconnection networks
should be done from an end-to-end perspective that factors in an understanding
of network traffic characteristics. It should include links, interfaces, and
traffic demands at the injection and reception endpoints of the network as much
as it does topology, switches, and links within the network fabric. Although
some rather straightforward performance models have been proposed in the past,
many do not consider well the end-to-end and traffic aspects; others are not
well suited to enable the network designer to reason about the network
tradeoffs being made. This talk will attempt to formulate a very simple, yet
general, performance model that provides an upper bound for end-to-end network
throughput and allows the designer to reason about various network design
tradeoffs and traffic behavior. It is based on the intuitive notion of modeling
the network as a "pipe" through which packets are transported end-to-end and
determining where the narrowest section or "bottlenecking point" occurs within
that pipe. Network topology, deadlock-free routing, arbitration, switching, and
flow control are all considered. Application of the model to a few case
examples (e.g., the Blue Gene/L and Cell Broadband Engine) are given, along
with some empirical validation via cycle-accurate simulation.
Timothy Pinkston is a Professor in the
EE-Systems Department of the Viterbi School of Engineering at the University
of Southern California, where he heads the SMART Interconnects research group.
Since January of 2006, Dr. Pinkston has been serving a 2-year term as a Program
Manager in the CCF Division of the CISE Directorate at the National Science
Foundation (NSF), where he directs research funding in the Computer System
Architecture area of the Computing Processes and Artifacts (CPA) Cluster.
His research interests include the development of deadlock-free adaptive
routing techniques and on-chip network and router architectures for achieving
high-performance communication in multicore and multiprocessor computer
systems. Most recently, Dr. Pinkston co-authored a book chapter with Professor
José Duato entitled, "Interconnection Networks" which appears
as Appendix E in the soon-to-be-released 4th edition of the best-selling text
Computer Architecture: A Quantitative Approach,
by John Hennessy and David Patterson.
|