The architectural landscape of high-performance computing stretches from superscalar uniprocessor to explicitly parallel systems, to dedicated hardware implementations of algorithms. Single-purpose hardware can achieve the highest performance and uniprocessors can be the most programmable. Between these extremes, programmable and reconfigurable architectures provide a wide range of choice in flexibility, programmability, computational density, and performance. The UCSC Kestrel parallel processor strives to attain single-purpose performance while maintaining user programmability. Kestrel is a single-instruction stream, multiple-data stream (SIMD) parallel processor with a 512-element linear array of 8-bit processing elements. The system design focuses on efficient high-throughput DNA and protein sequence analysis, but its programmability enables high performance on computational chemistry, image processing, machine learning, and other applications. The Kestrel system has had unexpected longevity in its utility due to a careful design and analysis process. Experience with the system leads to the conclusion that programmable SIMD architectures can excel in both programmability and performance. This work presents the architecture, implementation, applications, and observations of the Kestrel project at the University of California at Santa Cruz.

MOTIVATION: Sequence alignment is the problem of finding the optimal character-by-character correspondence between two sequences. It can be readily solved in O(n2) time and O(n2) space on a serial machine, or in O(n) time with O(n) space per O(n) processing elements on a parallel machine. Hirschberg's divide-and-conquer approach for finding the single best path reduces space use by a factor of n while inducing only a small constant slowdown to the serial version. RESULTS: This paper presents a family of methods for computing sequence alignments with reduced memory that are well suited to serial or parallel implementation. Unlike the divide-and-conquer approach, they can be used in the forward-backward (Baum-Welch) training of linear hidden Markov models, and they avoid data-dependent repartitioning, making them easier to parallelize. The algorithms feature, for an arbitrary integer L, a factor proportional to L slowdown in exchange for reducing space requirement from O(n2) to O(n1 square root of n). A single best path member of this algorithm family matches the quadratic time and linear space of the divide-and-conquer algorithm. Experimentally, the O(n1.5)-space member of the family is 15-40% faster than the O(n)-space divide-and-conquer algorithm.

Kestrel is a programmable linear systolic array processor designed for sequence analysis. Among other features, Kestrel includes an 8-bit word, a single-cycle add-and-minimize instruction, and efficient communication using systolic shared registers. This paper describes Kestrel's functional units in detail, and examines each of their effects on system performance. With prototypes currently in progress, we expect to complete a full Kestrel array, with between 512 and 1024 processing elements, by 1997.