Techniques for Co-Design of Optically-Connected Embedded Multiprocessors
Abstract
Several trends in technology have important implications for future digital signal processing (DSP) systems. By the year 2010, integrated circuit technology will allow 800 million transistors on a single chip. Already, manufacturers are placing multiple DSP cores on a single chip. Multiprocessor systems will become increasingly important in the future. A significant challenge is to develop software and compiler techniques to effectively exploit multiple processors. Signal and image processing algorithms are among those applications that can benefit from multiprocessor systems. Optics provides unique advantages and opportunities for designers of embedded multiprocessor systems, including the ability to construct highly connected and irregular networks that are streamlined for particular applications. Using these networks, it is possible to implement application mappings that allow flexible, low-hop communication patterns between processors. This has advantages for reduced system latency and power. Such optically connected multiprocessors are particularly promising for embedded DSP applications, which are highly parallel, and typically have tight constraints on latency and power consumption. Several groups have demonstrated optically-connected multiprocessor systems. However, comparatively little work has been done to develop compiler technology and automated mapping tools to take advantage of these systems. This work addresses the co-design of interconnect topologies and application mappings for DSP systems on optically connected multiprocessors. We demonstrate that existing DSP scheduling algorithms will deadlock for arbitrarily-connected networks, or when communication is restricted to a limited number of hops. We show that these low-hop communication schedules produce low power and low latency mappings.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 01, 2002
- Accession Number
- ADA483873
Entities
People
- Neal Bambha
- Shuvra Bhattacharyya
Organizations
- United States Army Research Laboratory