提升數據中心網路之用戶經驗服務品質

Abstract

近年來，由於諸如網頁搜尋、社群網路等新類型雲端服務的崛起，資料中心已經變得極為重要。作為一個不斷成長、且高收益的重要運算平台，資料中心必須要能夠滿足使用者的服務品質需求。在本論文中，我們將焦點專注於處理使用者在使用資料中心服務時，最為關心的因素中的兩項：服務延遲與回應資料品質。 由於頻寬是資料中心服務效能的一大重要瓶頸，因此在本論文中將專注於解決資料中心的頻寬議題。這議題可分為兩部分: 資料中心的內部頻寬與對外頻寬。在增加內部可得頻寬的方面，已經有許多優秀的方案被提出，但最終內部頻寬仍是有限的。因此我們專注於如何在有限的頻寬下，提供使用者更好的服務品質。為了達成此目標，我們提出一重要性覺察之資料遞送控制協定，藉此在有限的可得頻寬下，提供使用者較高的品質的回應資料。 而在資料中心對外頻寬的方面，過去主要依賴於內容傳遞網路(CDN)，或以多個分散的資料中心來降低服務延遲，但這樣的方式也會帶來極高的額外開支，並且頻寬依然會是有限的。在此議題上，我們提出了用便宜的儲存資源作為交換，利用使用者端未使用的頻寬與儲存空間，來減低資料中心頻寬需求，並在低成本的狀況下，提供低延遲的資料遞送服務。 透過最佳化內部頻寬的使用，與增加用於遞送靜態資料的外部頻寬來源，本論文提出解決資料中心頻寬議題的完整方式。實驗結果也顯示出，本論文的研究將可有效提升回應資料品質，並在服務延遲限制下，有效減低資料中心的頻寬負擔。

In recent years, data centers have played a critical role for fulfilling the requirements of new rapidly emerging services such as cloud storage, web search, online social network, and so on. Data centers, as critical computing platforms for ever-growing, high-revenue, must fulfill the requirements of those services for users. In this thesis, we aim to cope with two of the most concerned factors of users: the latency of services and the quality of results. In this thesis, we focus on one of the major obstacles of data center performance, the bandwidth bottleneck. The bandwidth bottleneck can be divided into two aspects, i.e., the internal bandwidth and the external bandwidth. We note that there are a many different approaches to improve the available internal bandwidth. However, the internal bandwidth is still a limited resource. Hence, we focus on providing better service quality, i.e., better quality of results, under the constraint of bandwidth. To achieve this goal, we propose a importance-aware transmission control protocol to provide better quality results before deadline. In contrast, the external bandwidth bottleneck is mostly relieved by content distribution network (CDN). However, the rental cost of CDN is high and the bandwidth is still limited by how much paid. Therefore, we exploit unused user-side bandwidth and storage to diminish the bandwidth requirement of data center. That is, we can provide low-latency data delivery service at low cost. The experiment results shows that the strategies on dealing with internal and external bandwidth bottleneck can effectively improve the quality of results and provide low-latency delivery service to users at low cost.