Real-Time Thread Level Load Sharing on Non-MMU Multi-Core Platforms

異質多核心平台在嵌入式系統的領域裡面被廣泛的應用著，例如行動通訊設備或者娛樂系統等等。針對異質多核心平台的特性，許多研究人員試著將管線排程技術(pipelined schedule technology)移植到這個平台上以增進此平台的系統效能。然而，靜態的管線排程演算法以及對執行時間的估計並無法滿足系統實際執行時的需求。因此，本篇論文提出了一個多核心工作負載共享策略(MCLSP)來解決這個問題。 而在最後的效能評估結果中，我們可以證明利用這個策，系統將可有效的降低具有即時性工作的最後期限缺失率(deadline miss rate)以及有效的平衡各核心間的工作負載。 此外，本篇論文還提出了一個工作負載搬移機制(thread migration mechanism)來實現在上述異質多核心平台中工作負載共享 的目的，此機制並解決了工作負載搬移時可能發生的資料不一致的 問題。Heterogeneous multi-core platforms are now well accepted for designing embedded systems including mobile devices and entertainment system. Many researchers proposed to execute the computation tasks in pipelined manner so as to improve system performance. However, the assumption on worst case execution time and pre-defined schedule make it not feasible to apply pipelined scheduling on open and complex embedded systems. In this thesis, we propose a multi-core load sharing policy, MCLSP, to solve this problem. The performance evaluation results prove the proposed MCLSP could reduce deadline miss rate of real-time tasks and balance the utilization rate of the processing cores as well. After MCLSP decides the target migration workload and the timing of load sharing, a workload migration mechanism should be applied in the system. Therefore, in this thesis, we also propose a workload migration mechanism for the heterogeneous multi-core platforms. And this mechanism can handle the data consistency problem on non-MMU multi-core architecture of this platform