在多核心架構上考慮效能與功耗問題的系統設計方法

Abstract

即使多核心系統已成為一般及嵌入式運算系統之主流，系統功耗及記憶體瓶頸仍然 限制了系統中的核心擴展性(Scalability)及平行性(Concurrency)。多核心性統同時也須考 慮溫度及電池之問題。執行序之間的親和度(Affinity)也限制了系統資源的使用率而產 生暗石夕晶(Dark Silicon)。

區域性(Locality)及附載平衡(Load Balance)為兩個和多心系統效能有關的重要的設 計問題。然而，針對單一問題而做出的最佳化可能導致另一問題更加嚴重，反之亦然。 而這兩個高度相關的問題需要一個協同的最佳化設計方法。而且，動態電壓及頻率 (dynamic voltage and frequency scaling)調整及功耗管理的技巧可能更進一步動態的調 整系統中處理器的運算能力並可能導致附載不平衡的問題。此計晝將為高平行度應用 發展高效率的區域性(Locality)及附載平衡(Load Balance)之最佳化方法。

在本計晝中，有三個關於功耗管理(Power Management)的問題將被考慮，分別是能 量消耗、溫度及效能。降低能量消耗或是溫度會造成一些效能上的損失，包含比較低 的運算速度，以及開關處理器造成的時間延遲問題。在這個計晝中，我們將會把此一 問題與負載平衡(Load Balancing)的問題一起做最佳化，以達到更短的整體運算時間。 由於本計晝著重於多核心系統架構(Many-Core Architecture)，因此我們將會設計線上 (online)且可擴展(scalable)的功耗管理政策。

儘管如此，目前的設計環境仍然缺少一個有效的方法論以提供大量的設計可能性探 索。在這份計劃中，我們將會考慮在多核心系統中兩個重要的議題，也就是速度及功 耗。此計晝預計的流程如下：第一年我們將會考慮靜態的區域性最佳化並準備基礎的 功耗管理政策，第二年則將發展考慮附載平衡的動態執行緒管理方法及階層式功耗管 理技術，而最後一年我們將在特殊系統架構下進行效能與功耗的同時最佳化 (Co-Optimization)。整合三年的成果後，我們將提出一個適用於多核心架構下跨層級且 多目標(Cross-Layer Multi-Objective)的設計流程。

As multiprocessors become the mainstream in both mainframe and embedded systems, the scalability and the concurrency are still limited by power wall and memory wall. The thermal issues and the limited capacity of batteries should be considered in multiprocessor systems. The affinity on different threads also restricts the resources from fully utilized; part of the processors become dark silicon.

Locality and load balance are two decisive design issues to exploit the superior performance of many-core architectures. However, optimizing only for one issue could deteriorate the potential benefit of the other. The highly correlated impacts of these two factors require a cooperative optimization. Furthermore, the dynamic voltage and frequency scaling and the other power management techniques would even change the computing ability of cores on the fly, which can also lead to load unbalance situation. This proposal will develop highly efficient optimization algorithm to enhance the locality and load balance of massive parallel applications.

This project considers three issues of power management for multiprocessor systems: energy, thermal and performance penalty. Reducing energy consumption or lowering temperature may cause performance penalties, including frequency degradation and wakeup delay. In this project, it is co-optimized with the load-balance problem to obtain shorter latency. Since this project focuses on many-core architectures, we will design scalable and online policies.

In summary, the current design environment lacks an effective methodology to explore the huge design space. In this proposal, we will consider the performance and power issues in multiprocessor systems. The research process of this project will be scheduled as following: we will consider static locality-aware optimization and prepare the fundamental policies for power management in the first year, develop dynamic load-balance-aware thread management and hierarchical power management techniques in the second year, and architecture-aware co-optimized runtime management in the last year. In conclusion, we will combine the results of three years to form a cross-layer multi-objective methodology for many-core architecture.