Investigation of parallel programming on heterogeneous multiprocessors
Håvard Espeland · NORA - Norwegian Open Research Archives · 2008
Multi-core processors have become ordinary in modern commodity computers. Computationally intensive applications, like video processing, that previously only ran on specialized hardware, are now common on home computers. However, the demand for more computing power is ever-increasing, and with the introduction of high definition video, more performance is desired. As an alternative to having multiple identical processor cores, heterogeneous multiprocessors have cores with different capabilities. This allows tasks to be processed on simple cores with specialized functionality. The simplicity furthers low power consumption, small die usage, and low price. Dealing with heterogeneous cores increases the complexity of writing programs for the architecture. The reasons for this includes different capabilities of the cores, and some heterogeneous architectures do not have shared memory. Without shared memory, accessing main memory requires explicit transfers to local memory. In this thesis, we consider two architectures, the STI Cell/B.E. and Intel IXP2400, and evaluate parallelization strategies and performance for real-world problems. Our tests show promising throughput for some applications, and we propose a scheme for offloading computationally intensive parts of an existing application.