Motivation
Many important applications with massive memory and/or computational requirements can be sped
up on a multiprocessor system. However, today's parallel machines suffer in varying degrees from
poor programmability, inadequate network performance, and limited scalability.
There is a need for programmable high-bandwidth systems which are scalable to one million
processing nodes and beyond.
To achieve this goal it is necessary to reevaluate system design from top to bottom. Monolithic
processors which devote millions of transistors to running a sequential stream of instructions
must give way to simpler and more efficient processing nodes which allow multiprocessing
on a single die and provide greater overall performance. A high bandwidth network is required
for running communications-intensive programs with poor or unpredictable locality. The operating
system must manage system resources securely without hindering performance. Finally, a programming
language and compiler are needed which allow programmers to express parallel constructs in a
natural manner.
The Aries research group consists of a number of integrated projects which will be brought
together in the design and implementation of a complete system. Both the design process and
the final product will be instrumental in understanding the challenges associated with
developping a successful massively parallel machine.
Research Goals
- Understand the language, compiler, operating system and architectural requirements which
must be satisfied in order to develop a highly programmable parallel system
- Investigate techniques for the management of huge data sets
- Develop an area-efficient processor architecture
- Explore architectural features enabled by the integration of logic and memory
Architectural Goals
- Efficient general purpose processor (1 GFLOP/5 million transistors)
- High performance network
- High bandwidth (4GB/s to each node)
- Low latency (10ns point to point for a 1000 node system)
- Scalability (up to 1,000,000 nodes)
This work is supported by DARPA/AFOSR contract number F306029810172
Site design by J.P. Grossman