Document Type

Technical Report

Department

Computer Science and Engineering

Publication Date

2006-01-01

Filename

wucse-2006-19.pdf

Technical Report Number

WUCSE-2006-19

Abstract

Profile Hidden Markov models are highly expressive representations of functional units, or motifs, conserved across protein sequences. Profile-HMM search is a powerful computational technique that is used to annotate new sequences by identifying occurrences of known motifs in them. With the exponential growth of protein databases, there is an increasing demand for acceleration of such techniques. We describe an accelerator for the Viterbi algorithm using a two-stage pipelined design in which the first stage is implemented in parallel reconfigurable hardware for greater speedup. To this end, we identify algorithmic modifications that expose a high level of parallelism and characterize their impact on the accuracy and performance relative to a standard software implementation. We develop a performance model to evaluate any accelerator design and propose two alternative architectures that recover the accuracy lost by a basic architecture. We compare the performance of the two architectures to show that speedups of up to 3 orders of magnitude may be achieved. We also investigate the use of the Forward algorithm in the first pipeline stage of the accelerator using floating-point arithmetic and report its accuracy and performance.

Comments

Permanent URL: http://dx.doi.org/10.7936/K7WM1BNJ

Download

Share

COinS