Topics for Distinguished Visitors Program
(Jie Wu)
Parallel Processing: Past, Present, and Future 1
Parallel computing
is one of the most exciting technologies to achieve prominence since the
invention of electronic computers. But the departures from the Von Neumann
computational model create many questions and problems. What kind of applications
have sufficient parallelism? What will be the nature of the processing
elements, and how will they communicate with each other? What will the
new computational model be? What will the software be like? This talk
focuses on the total design problem involved in parallel processing. A
brief history of parallel computing is presented first, followed by a
list of major issues in the design of a parallel system. Three major design
approaches are discussed: control (task) parallelism, data parallelism,
and pipelining. Finally, we provide our own view of the future of parallel
processing.
Fault-Tolerant Communication in Point-to-Point Networks
One of
the challenges in fault-tolerant communication is to achieve certain objectives
without excessive cost in collecting and maintaining fault information.
We propose an approach that achieves optimal routing with limited amount
of global fault information. This approach is illustrated by the concept
of a safety level (vector) that can be applied to hypercube- and mesh-connected
systems. In this approach each node is associated with a safety level,
which is an approximate measure of the number and distribution of faulty
nodes in the neighborhood. The safety level can be used to determine routing
feasibility as well as to provide necessary information to facilitate
optimal routing. This talk addresses related issues and possible future
directions in this area.
Optimal Routing in Mesh-Connected Multicomputers: Challenges
and Solutions
Mesh-connected
topology is one of the most thoroughly investigated network topologies
for multicomputer systems. It is important due to its simple structure
and its good performance in practice. We discuss several optimal routing
problems which are important for system performance which is dependent
on the end-to-end cost of communication mechanisms. We then focus on two
optimal routing examples, one for one-to-one routing (also called unicasting),
and the other for one-to-all routing (also called broadcasting). Applicability
of these approaches and open directions of research in this area are also
presented.
Since July 2003
|