Efficient parallel solutions to the integral knapsack problem on current chip-multiprocessor systems

摘要

The emergence of chip-multiprocessor systems has dramatically increased the performance potential of computer systems. However, harnessing the full potential of these systems depends largely on the effectiveness of system software, such as compilers, in exploiting the on-chip parallelism. Additionally, since the amount of parallelism extracted by a compiler is directly influenced by the selection of the algorithm, algorithmic choice also plays a critical role in achieving a high fraction of peak performance. Hence, in the era of multicore computing, it is imperative that we re-evaluate and rethink algorithms for key problem domains. This paper investigates the impact of algorithmic choice on the performance of parallel implementations of the integral knapsack problem on multicore architectures. The study considers two classes of algorithms and several algorithmic variants and evaluates each implementation based on a variety of performance metrics including data locality and sharing, granularity of parallelism and scalability. The paper presents experimental results that show how each performance factor is affected by the selection of algorithm, changes in the input data-set and variations in architectural characteristics such as cache capacity and degree of cache sharing.