基於耗能與時間效能的異質系統工作分配管理

Abstract

在異質計算的系統下，如何有效的將工作分配至現有的硬體資源來達到最短的執行時間一直是一個很關鍵的議題，本論文不僅討論在提高運算速度下的工作分配，更多了整體功耗的考量，且探討加入現今異質系統架構(Heterogeneous System Architecture)的特性:異質統一記憶體存取(Heterogeneous Uniform Memory Access, hUMA)與共享虛擬記憶體(Shared Virtual Memory, SVM)，在達到的CPU與GPU的共享記憶體存取後(透過此技術在異質系統下將不須傳統的跨裝置的資料傳送)，工作分配產生的變化。 本研究使用AMD的Kaveri 7850K APU (Accelerated Processing Unit)，搭配OpenCL使單一kernel的工作以比例形式將工作物件(workitem)分配給CPU與GPU執行，並比較以時間效能為目標與以最低能量延遲乘積(energy delay product, EDP)為目標時的最佳分配比例的不同，並使用應用程式的動態特性與靜態特性的建立預測模型，並探討在不同目標(時間或EDP)下最佳工作分配比例與應用程式特性相關性的變化，之後透過OpenCL 2.0 API實現共享記憶體存取技術，並分析其最佳工作分配比例的變化。

On the heterogeneous system, how to achieve the shortest execution time by using task partition on the existing hardware resources is a critical issue. In this paper, we not only consider the problem of partitioning works to increase speed but also observe total power issue. In addition, we investigate the problem of partitioning works in the HSA (Heterogeneous System Architecture) features: hUMA(Heterogeneous Uniform Memory Access) and SVM (Shared Virtual Memory). In this new situation, the platform don’t need the traditional data transfer between CPU and GPU anymore. In this paper we adopt AMD Kaveri 7850K APU(Accelerated Processing Unit) platform. We leverage a ratio value to partition an OpenCL kernel mapping on CPU device and GPU device. The optimal ratio of the best speed performance would compare with the optimal ratio of the lowest EDP (energy delay product). We build a prediction model based on program static features and runtime features, and discuss the difference of correlation coefficient between performance and EDP target. Finally, we analyze the variation of optimal partition ratio in the situation of SVM technique with OpenCL 2.0 API.