在開發初期,軟體和硬體開發者對於多核心架構需要一個可以快速評估效能的方式。 在這篇論文中,我們提出了一個的比執行驅動模擬(execution-driven simulation)和追蹤驅動模擬(trace-driven simulation)更具效率的多核心效能模型的模擬架構,藉由擷取機器參數和分析應用程式的效能資訊去實現機器相關的分析和建構抽象的效能模型。為了模擬多核心系統中應用程式的資源競爭行為,我們建構了一個記憶體競爭情況的統計模型去預估發生資源競爭所產生的延遲。藉由提高模擬層級,可以有效地降低模型複雜度與所需的儲存空間。我們選擇了數個benchmark評估提出的機制,在與執行驅動模擬(execution-driven simulation)對於循序與多線程執行的benchmark的比較中,我們提出的機制可以分別達到2407倍和28倍的效能改進;在與追蹤驅動模擬(trace-driven simulation)的比較中,我們提升了至少3800倍的效能。
Software and hardware developers need to get a estimation of performance quickly on a multicore architecture during the early design phase. We proposed a performance modeling simulation framework that achieves higher speedup than both execution-driven simulation and trace-driven simulation by extracting the machine parameters to carry out machine-dependent analysis and analyzing the application performance information to be an abstract model. For modeling the resource contention condition in the multicore system, we also build the statistical memory contention model to predict the contention delay time. By raising simulation level to higher levels of abstraction, modeling complexity is lower and the input of storage is smaller. The proposed simulator achieves average 2407x and 28x higher than emph{QEMU} in sequential and multithreaded applications individually. Our simulator achieves at least about 3800x speedup against trace-driven simulator, for the benchmark programs in the case study.