◾ Multicore Architectures

Abstract

In this dissertation, we address the challenges of performance scaling for bioinformatics applications on multicore architectures. In particular, we focus on sequence alignment, one of the fundamental tasks in bioinformatics. Due to the exponential growth of biological databases and the computational complexity of the algorithms used, high performance computing systems are required. Recently, computer architecture has shifted towards the multicore paradigm in an attempt to sustain the performance scalability that single-core processors did provide in the past. Although multicore architectures have the potential of exploiting the task-level and data-level parallelism found in bioinformatics workloads, efficiently harnessing systems with hundreds of cores requires deep understanding of the applications and the architecture specifics. This thesis presents a study of two sequence alignment applications that are modeled, characterized, mapped and optimized targeting two multicore architectures. More precisely, we use the Cell BE and the SARC architectures, the latter developed within the project this thesis was part of. The targeted applications, i.e., HMMER and ClustalW, are used for pairwise alignment and multiple sequence alignment. We first propose an analytical model to predict the performance of applications parallelized under the master-worker scheme. We use HMMER and the Cell BE processor for our experimental case study and for validation of our model. Results show the high accuracy of the model and the scaling behavior of the application phases. Next we investigate the optimal mapping of ClustalW on the Cell BE, identify a number of limitations in the architecture and propose few instruction-set extensions to accelerate the main ClustalW kernel. Last, we study ClustalW and HMMER scalability on the SARC architecture with up to one thousand cores. We also investigate the impact of different input types on ClustalW performance.