Scalability of Bioinformatics Applications for Multicore Architectures

Exponential growth in biological sequence data combined with the computationally intensive nature of bioinformatics applications results in a continuously rising demand for processing power. Microprocessor complexity and, more importantly, computational capability increases as well, through transistor budgets that grow in line with Moore's law. However, limits in power consumption, frequency scaling and memory technology cause single threaded performance improvements to stagnate. The result is a paradigm shift to parallel architectures, an example being the state-of-the-art Cell Broadband Engine. In this thesis, suitability of this architecture is examined for HMMER, a bioinformatics application that identifies similarities between protein sequences and a protein family model. Qualitative and quantitative analysis is performed to reveal its scaling behavior and potential bottlenecks. Inspection of the program structure shows that the parallelization strategy imposes limits on scaling ability. Based on function profiling, a model for application performance is proposed that is accurate within 2%. From the model, the optimal PPE/SPE ratio is derived for different workloads. For typical workloads, the PPE can supply nine SPEs with jobs. The TaskSim simulator, whose phase-based simulations are accurate within 2%, is used to validate the model's predictions and to demonstrate that scaling behavior is mostly determined by the buffering of jobs.Microelectronics & Computer EngineeringElectrical Engineering, Mathematics and Computer Scienc