Microarchitecture Characteristics and Implications of Alignment of Multiple Bioinformatics Sequences

With the growth of bioinformatics and computational biology industry, multiple sequence alignment (MSA) applications have become an important emerging workload. In spite of the large amount of recent attention given to the MSA software design, there has been little quantitative understanding of the performance of such applications on modern microprocessors and systems. In this paper, we analyze performance and characteristics of MSA software from the perspective of processor microarchitecture. We use twelve popular MSA programs employing a wide variety of alignment approaches. The basic workload characteristics and the efficiencies of various microarchitecture features, such as trace cache, out-of-order execution, caching, branch prediction, speculative execution and phase behavior are examined and analyzed on the Pentium 4 microarchitecture. Our major observations of this work are: (1) Instruction footprints of MSA programs are typically small and can fit in the L1 instruction cache. The trace cache shows high percentage of time in deliver mode. (2) Loads and stores account for the 60% of dynamic instructions executed. This indicates that further improving memory bandwidth will be beneficial to the performance of MSA software. (3) Prefetching and large L2 cache can efficiently handle the working sets of a majority of the studied benchmarks. Nevertheless, MSA software using exact alignment method yields poor cache performance. The data TLB behavior largely depends on the alignment methods used. (4) The studied MSA applications show large variation in dynamic branch frequency and mix. The indirect branches, calls, and returns can be predicted with high accuracy. The overall branch misprediction rates exceed 5% on half of the examined benchmarks. (5) The IPC of the studied benchmarks ranges from 0.15 to 0.93. Overall, the processor speculatively executes 27% more instructions than it retires.