行動裝置社交型虛擬實境應用程式在雲端系統中高效能的整體網路資源排程

Abstract

行動裝置雲端計算系統將行動裝置應用程式上的工作負載卸載到雲 端系統。本篇碩士論文著重於卸載具有高通信與計算比的行動裝置社 交型虛擬實境應用程式，這使得網路資源成為雲端系統中的瓶頸。而 我們的目標是在提供服務時降低網路硬體資源需求同時保證服務品質， 並且在無需添加更多硬體資源的情況下最大化社交型虛擬實境實例被 容納的數量。 在沒有事先知道社交型虛擬實境實例未來需求的情況下，很難適當 地減少被啟用的硬體機器數量，以及在沒有額外硬體資源的情況下， 也很難增加社交型虛擬實境實例的數量。因此，我們提出了一個新的 排程演算法，Profiling and Partition Scheduling （PPS），以剖析應用程 式的網路需求，再利用統計分析的方式來預測未來的需求進而減少啟 用機器的數量，同時保證服務品質的違規率處於極小值。接著，我們 利用社交型虛擬實境應用程式的特性，提出一個分治技術以增加同時 容納實例的數量。 我們實做一個測試用雲端系統以評估我們的 PPS 排程演算法並使用 模擬評測和實際實驗。評測顯現出我們將網絡使用率提高到 72%、減 少所需機器的平均數量，並保證服務品質違規率低於 0.13% 以下。另 外，我們在模擬中減少了被拒絕服務的實例數量高達 25%，而實際實 驗中高達 13%，我們以此證明我們可以在不額外增加硬體資源的限制 下，增加社交型虛擬實境實例的數量。

Mobile cloud computing offloads workload from mobile application into cloud system. This paper focuses on offloading a mobile social VR application with high communication-to-computation ratio, which makes the network a bottleneck in this application in the cloud. Our goal is to reduce the network hardware requirement while providing Service Level Objective (SLO) guarantee, and to maximize the number of accommodated social VR instances without adding more hardware resources. It is difficult to reduce the amount of active hardware without knowing the future demand of a social VR application in advance, and it is also difficult to increase the number of accommodated social VR instances without extra hardware resources. Thus, we propose a new scheduling, Profiling and Partition Scheduling (PPS), to profile the application and use statistical analysis to predict future demand to reduce the number of active machines while guaranteeing a small SLO violation rate. Then, we exploit the property of social VR application to propose a partition technique to increase the number of simultaneously accommodated instances. We implement a cloud testbed to evaluate our PPS scheduling algorithm and conduct both trace-driven simulation and real-world experiment. The evaluation shows that we increase the network usage rate up to 72%, decrease the average number of required machines, and guarantee an SLO violation rate under 0.13%. In addition, we decrease the number of rejected instance by 25% in the trace simulation and 13% in running a real-world application iv on the testbed. That is, we demonstrate that we can increase the number of accommodated social VR instances without adding more hardware resources.