軟體定義網路之自適性串流技術研究

Abstract

視訊串流可視為現今最重要的即時性(real time)網路服務之一，它改變了世界上多數人的收視習慣。為了最小化暫存空間需求並提供最佳的使用者體驗，因此了所謂自適性串流(adaptive bitrate streaming, ABS)的新興視訊串流技術。這種技術藉由持續評估網路頻寬與流量狀態，以及中央處理單元(central processing unit, CPU)與記憶體(memory)狀態，來達到每個用戶端都盡量分配到符合當前網路狀況的視訊位元率(video bitrate)的目標。 雖然自適性串流裝置可藉由持續不斷地觀測網路狀態並評估可用頻寬資訊，來動態地選擇適當的視訊位元率。然而，傳統的自適性串流技術賦予每個裝置有權決定需要多少頻寬資源，造成各裝置端個別的視其端點網路速度選擇對其最有利的視訊位元率。這種分散式的決策，會因而出現某些裝置高估了自身所需求的頻寬資源，而造成頻寬分配不公的現象。譬如一些裝置分配到高於它們實際需求的頻寬資源，而有部分的裝置卻被分配到低於他們所需求的頻寬。何況高估可用頻寬資源可能會在下次迭代時造成其他裝置低估其可用頻寬資源，最終帶來不穩定的視訊位元率切換和較差的使用者體驗(quality of experience, QoE)。 為了要評估視訊串流服務的使用者體驗，我們通常使用影像品質(video quality)作為度量的基準。因此，要達成最佳的使用者體驗就等於要達成最佳的公平影像品質。我們將使用者體驗最佳化的問題定義為一個最大化最低公平性的問題(maximum minimum fairness, MMF)。我們提出了三種方法以處理MMF問題，其概念為在所有可以被分配的頻寬(影像位元率)以及其所對應畫質中找到能讓收視畫質最差的用戶之影像畫質的公平性被最大化的分配法。另一方面，為了實現MMF的解決方案，應有一資源分配者可收集到所有裝置相關的網路狀況資訊，而這樣的網路須是可由軟體控制的架構。軟體定義(控制)網路(software defined network, SDN)有一中央管控的平台，能夠持續監管網路整體運行狀態，並透過蒐集相關的資訊來智慧化地管理網路流量。網路中的自適性串流裝置無須自行決定影像位元率，只需要傳遞目前網路流量與裝置儲存資訊給軟體定義網路的控制端再由其分配頻寬資源。我們所提出的即時頻寬資源分配的解決方案即是基於此種軟體定義網路架構，來最大化最低的公平使用者體驗。我們同時以電腦模擬與硬體實作來驗證所提出方法的可行性與效能，並確認所提出的三個演算法皆能達成使用者體驗的公平化，亦即最大化最低公平性。

Video streaming is one of the most popular real-time Internet services which has changed the viewing behaviors of millions of people around the world. A novel streaming technique called adaptive bitrate streaming (ABS) that allocates appropriate video bitrates based on the current overall network capacity (especially CPU and memory) and traffic condition can minimize the buffering requirement and provide satisfactory user experiences to all viewers. By monitoring the network condition and evaluating the available bandwidth continuously, an ABS client adaptively selects an appropriate video bitrate. However, conventional ABS techniques endow each individual client with full authority to determine the desired bandwidth. Each client thus unilaterally observes the network traffic and makes a video bitrate decision which serves its demand best. The distributed decisions entail unfairness in bandwidth allocation as ABS clients tend to overestimate the required bandwidth. As a result, some may be allocated more bandwidth than actually needed while others receive less than they really need. Furthermore, an overestimation may cause an underestimation in the next iteration, ultimately bringing the clients with unstable video bitrates and poor quality of experience (QoE). To evaluate QoE in video streaming services, video quality is often used as a proper metric. Thus, optimizing QoE is equivalent to optimizing the “video quality” fairness. We formulate the QoE optimization problem as a maximum minimum fairness (MMF) problem. It essentially searches for a candidate bandwidth allocation (in bitrate) and the corresponded video quality for all the clients so that the worst client video quality fairness is maximized. Three schemes are proposed to solve the MMF problem. On the other hand, for these solutions to be implementable, the information regarding the network conditions of all clients should be available to a resource allocation agent. It is clear this is realizable only if the network in question has a software defined architecture. A software defined network (SDN) has a centralized controller platform which continuously monitors the overall network condition and collects related information to manage flow control for intelligent networking. The ABS clients need not to make bitrate decisions but simply forward the observed network traffic and storage status to the SDN controller. Based on the SDN architecture, our numerical solution gives a resource allocation policy for ABS clients to achieve the mini-max QoE fairness in real-time. Both computer simulation and hardware implementation results are provided to verify the feasibility and efficiency of the proposed methods. We find that all three algorithms achieve the same QoE fairness (i.e., MMF).