隨著行動裝置、邊緣運算及高效能運算（HPC）平台的蓬勃發展，現代運算需求與日俱增。異質多核心系統（Heterogeneous Multicore Systems）因為能整合不同特性的處理核心進行協同運作，提供卓越的運算能力與資源利用率，已成為當前主流的硬體架構。然而，異質架構在提升效能的同時，也大幅增加了任務調度（Task Scheduling）的複雜度。為配合複雜的運算要求，晶片製造廠力求晶片微縮，但是造成電晶體更脆弱，對系統可靠度（System Reliability）造成嚴重威脅，可能使壽命減短。在前面因素影響下，由於壽命變短，頻繁地汰換硬體，以及使用時耗能增加，導致大量碳排放（Carbon Emissions），致使全球暖化，因此碳排放亦成為現代系統架構設計中不可忽視的課題。在異質多核心系統上，如何透過優化排程最大化壽命以及盡量減少碳排放，已成為研究核心目標。研究主要會分成離線（offline）和線上（online）兩階段，先在離線時候，會以低碳排壽命長為目標，訓練一個模型，供線上的時候，可以指引蒙地卡羅樹搜尋的方向，往模型評估較好的候選靠，避免花時間將所有候選試過一遍才做決定，這種策略能對於任務的調度可以提供即時的答案，使得任務執行不用等待。實驗結果顯示，此方法有效地延長了系統的壽命（lifetime），並同時減少碳排放，與透過平衡老化以延長壽命的排程方法相比較，在壽命方面超越了最多至8.04倍，同時平均碳排放又減少了最多至17.54%。;With the rapid development of mobile devices， edge computing， and high-performance computing (HPC) platforms， modern computing demands continue to increase significantly. Heterogeneous multicore systems have become the predominant hardware architecture due to their ability to integrate processing cores with diverse characteristics to support flexible task execution， thereby improving computational capability and resource utilization. However， while enhancing performance， these heterogeneous architectures significantly increase the complexity of task scheduling. To meet increasingly demanding computational requirements， chip manufacturers continue to scale down semiconductor technology. Nevertheless， such technology scaling makes transistors more vulnerable， posing a serious threat to system reliability and potentially shortening system lifetime. Hence， frequent hardware replacement due to short lifetime and increased energy consumption during operation lead to substantial carbon emissions， which in turn aggravate global warming. Therefore， carbon emissions have become an important issue that cannot be ignored in modern system architecture design. In heterogeneous multicore systems， how to optimize task scheduling to maximize system lifetime while minimizing carbon emissions has become a key research objective. To address this issue， this thesis proposes a two-stage framework consisting of offline and online phases. In the offline phase， a model is trained with the objective of achieving low carbon emissions and long system lifetime. In the online phase， the trained model is used to guide the search direction of Monte Carlo Tree Search (MCTS) toward candidates that are evaluated as more promising by the model， instead of exhaustively exploring all possible candidates before making a decision. This strategy enables the scheduler to provide timely scheduling decisions， ensuring that task execution does not need to wait for a prolonged search. Experimental results demonstrate that the proposed method effectively extends system lifetime while simultaneously reducing carbon emissions. Compared with existing scheduling methods that extend system lifetime through balanced aging across cores， the proposed approach improves system lifetime by up to 8.04× while simultaneously reducing average carbon emissions by up to 17.54%.
