電機資訊學院: 資訊工程學研究所指導教授: 劉邦鋒; 吳真貞周儒成Chou, Ju-ChengJu-ChengChou2017-03-032018-07-052017-03-032018-07-052015http://ntur.lib.ntu.edu.tw//handle/246246/275407現今行動裝置越來越普及，而使用者對於能在行動裝置上獲得之體驗也有更多期待。因此，開發商嘗試於在行動裝置上進行硬體整合，像是導入傳統電腦上會使用的GPU處理器。但行動裝置上的運算能力還是普遍不足，在運算較複雜的工作還無法展現出足夠的效能。近年來許多研究利用雲端運算架構將運算工作移轉致遠端設備協助處理來得到較佳效能，但網路傳輸的成本在這個架構中有相當程度的影響，如何在不穩定的網路環境中進行移轉運算也成為一個重要的議題。 在本篇論文中，我們設計並實作了一OpenCL運算工作移轉架構，藉由利用降低網路傳輸的成本以及善用異質化運算的優勢來加速行動裝置上的OpenCL運算工作。我們利用機器學習機制決策出整體運算工作在行動裝置以及遠端運算設備的分配，讓兩端設備進行分散式運算以完成整體運算。通過這樣的分散式運算，我們可以善用兩端的運算資源以及減少資料的傳輸量以降低網路的傳輸成本。實驗結果指出，我們的決策機制可以找出合適的分配量，即使在動態環境中，我們的決策結果在效能上相較於最佳分配最多僅有8%的誤差。Mobile devices have become increasingly prevalent in recent years, leading the public to reassess their expectations in terms of user experience. As a result, industry has progressively turned to utilizing hardware components, such as GPUs (graphics processing units), which are traditionally present on computers. Despite their sophistication, mobile devices lack the capacity to execute resource-intensive tasks quickly and efficiently, and the state of the art is to leverage heterogeneous cloud resources to augment mobile devices. However, network transfer costs drastically limit the advantage of this approach. While performing computing tasks on a heterogeneous system is a well-studied area, how to offload workload onto a heterogeneous cloud in the presence of an unstable network remains an outstanding problem. This thesis presents the design and implementation of a workload offloading framework that transparently mitigates the network transfer cost and takes advantage of a heterogeneous resourcerich cloud for speeding up mobile devices’ OpenCL computations. Our approach uses machine learning mechanism to decide the workload partition and maps the processing elements to heterogeneous local and remote resources. By partitioning the tasks, our system increases the utilization of mobile and server resources while reducing the amount of data to transfer over the network. Our results show that adaptive partitioning can have a significant impact on the performance of benchmarks, even in a dynamic environment. The experiments result shows that the difference in performance between the optimal partition and the partition suggested by our model is less than 8%.587801 bytesapplication/pdf論文公開時間: 2015/7/29論文使用權限: 同意有償授權(權利金給回饋學校)OpenCLGPGPU行動雲端運算普及運算運算移轉異質化系統資源分配Mobile Cloud ComputingPervasive ComputingComputation OffloadingHeterogeneous SystemResource Mappingthesishttp://ntur.lib.ntu.edu.tw/bitstream/246246/275407/1/ntu-104-R02922028-1.pdf