Improving the scalability of multicore systems with a focus on H.264 video decoding

In pursuit of ever increasing performance, more and more processor architectures have become multicore processors. As clock frequency was no longer increasing rapidly and ILP techniques showed diminishing results, increasing the number of cores per chip was the natural choice. The transistor budget is still increasing and thus it is expected that within ten years chips can contain hundreds of high performance cores. Scaling the number of cores, however, does not necessarily translate into an equal scaling of performance. In this thesis, we propose several techniques to improve the performance scalability of multicore systems. With those techniques we address several key challenges of the multicore area. First, we investigate the effect of the power wall on future multicore architecture. Our model includes predictions of technology improvements, analysis of symmetric and asymmetric multicores, as well as the influence of Amdahl's Law. Second, we investigate the parallelization of the H.264 video decoding application, thereby addressing application scalability. Existing parallelization strategies are discussed and a novel strategy is proposed. Analysis shows that using the new parallelization strategy the amount of available parallelism is in the order of thousands. Several implementations of the strategy are discussed, which show the difficulty and the possibility of actually exploiting the available parallelism. Third, we propose an Application Specific Instruction Set (ASIP) processor for H.264 decoding, based on the Cell SPE. ASIPs are energy efficient and allow performance scaling in systems that are limited by the power budget. Finally, we propose hardware support for task management, of which the benefits are two-fold. First, it supports the SARC programming model, which is a task-based dataflow programming model based on StarSS. By providing hardware support for the most time-consuming part of the runtime system, it improves the scalability. Second, it reduces the parallelization overhead, such as synchronization, by providing fast hardware primitives