Characterizing and Accelerating Bioinformatics Workloads on Modern Microarchitectures
Characterizing and Accelerating Bioinformatics Workloads on Modern Microarchitectures
Files
Publication or External Link
Date
2007-04-25
Authors
Albayraktaroglu, Kursad
Advisor
Franklin, Manoj
Citation
DRUM DOI
Abstract
Bioinformatics, the use of computer techniques to analyze biological data, has been a particularly active research field in the last two decades. Advances in this field have contributed to the collection of enormous amounts of data, and the sheer amount of available data has started to overtake the processing capability possible with current computer systems. Clearly, computer architects need to have a better understanding of how bioinformatics applications work and what kind of architectural techniques could be used to accelerate these important scientific workloads on future processors. In this dissertation, we develop a bioinformatic benchmark suite and provide a detailed characterization of these applications in common use today from a computer architect's point of view. We analyze a wide range of detailed execution characteristics including instruction mix, IPC measurements, L1 and L2 cache misses on a real architecture; and proceed to analyze the workloads' memory access characteristics. We then concentrate on accelerating a particularly computationally intensive bioinformatics workload on the novel Cell Broadband Engine multiprocessor architecture. The HMMER workload is used for protein profile searching using hidden Markov models, and most of its execution time is spent running the Viterbi algorithm. We parallelize and partition the HMMER application to implement it on the Cell Broadband Engine. In order to run the Viterbi algorithm on the 256KB local stores of the Cell BE synergistic processing units (SPEs), we present a method to develop a fast SIMD implementation of the Viterbi algorithm that reduces the storage requirements significantly. Our HMMER implementation for the Cell BE architecture, Cell-HMMER, exploits the multiple levels of parallelism inherent in this application, and can run protein profile searches up to 27.98 times faster than a modern dual-core x86 microprocessor.