Supporting Multi-handoff Mechanism for Streaming Service in a Web Cluster with Content-Aware Request Distribution

[[abstract]]由於網際網路科技的爆炸性發展，影音串流服務已十分普及。YouTube、無名小站、土豆網和優酷網都是串流服務中最著名的多媒體網站。由於單一網路伺服器已無法處理在串流服務中相當沉重的工作量，因此叢集式伺服器透過分享負載到多台伺服器的方式已成為一個有效的解決方案。叢集式系統是由一台負載平衡伺服器(又稱為前端伺服器)和多台實體伺服器(又稱為後端伺服器)所組成，它可以彈性地增加或移除實體伺服器及網頁服務。因此，叢集式系統極具成本效益，且具備高度的擴充性與高效能等優點。 當串流服務運行時，傳統上一個串流連線是不能被移轉至其他伺服器直到整部影片都播放完畢。如果一部冗長的影片佔據了一台後端伺服器的硬體資源，它極有可能在叢集式系統中引發負載不平衡的情形。為了解決這個問題，在過去的研究中，我們已提出且實作RTSP Handoff機制[7]於LVS-CAD叢集式系統上，此機制能將一部影片在邏輯上分成數段固定的長度，且每一影片分段的播放是可由不同的後端伺服器來提供影片內容。然而，此機制並沒有彈性，它不能根據影片的長度來決定一部影片該分成幾段。如果一段已分段的影片仍然太長，它仍可能在整個叢集式系統中引發負載不平衡的情形。 在這篇論文中，我們進一步地提出了RTSP Multiple Handoff機制來改良RTSP Handoff機制，透過此改良的機制可依一部影片的長度動態地決定該分成幾段，即將播放一部影片的負載分別由不同的後端伺服器來接續服務。我們所提的機制是實作在Linux kernel 2.6.18版本的LVS-CAD叢集式系統上。在我們的實驗中，我們測試了在客戶端所要求播放的影片有短中長片不同比例的情況下，我們比較LVS-CAD使用RTSP Multiple Handoff機制或RTSP Handoff機制，以及LVS沒有使用這些機制的產出(throughput)及平均回應時間(average response time)。實驗結果顯示，在LVS-CAD 叢集式網頁伺服器上，使用RTSP Multiple Handoff機制將比使用RTSP Handoff機制在LARD/RC的需求分配機制及短中長片播放比例為2:1:1的情況下，可以獲得24.43%更佳的效能。同樣的在LVS-CAD使用LARD/RC需求分配機制但不同影片播放的比例之下，也能獲得比LVS使用WLC與WRR有36.34%至107.52%更佳的效能。實驗結果也顯示當LVS-CAD使用RTSP Multiple Handoff機制，可以比LVS沒有使用RTSP Multiple Handoff機制且使用WLC與WRR需求分配機制時減少49.03%至91.51%的平均反應時間。[[abstract]]As the explosive development of technologies in the Internet, streaming services have become very popular to end-users. YouTube, Wretch, Tudou, and Youku are the most well-known multimedia web sites providing streaming services. Because a single server cannot handle the heavy workload of streaming services, a web cluster is an effective solution by sharing loading among several servers. A web cluster is composed of a load balancer also called front-end server and several real servers also called back-end servers. A server can be added into or removed from the back-end servers dynamically. A web cluster has the advantages of cost effectiveness, scalability, and high performance. Traditionally, when a streaming service is running, the current streaming connection between a client and a server cannot be handed off to another server until the film is entirely played back. If a film is too lengthy and thus occupies the resource of a back-end server for a long time to serve this film, it may cause severe load imbalance in the whole web cluster. To solve this problem, in our previous work, the RTSP Handoff mechanism is proposed [7] and implemented on LVS-CAD [5] web cluster. This mechanism can logically partition a film into several sections which are served by different back-end servers. However, it does not have the flexibility to determine how many sections a film has to be divided. If a film section is too long, it may still cause load imbalance on the whole web cluster. In this thesis, we have proposed the RTSP Multiple Handoff mechanism to improve the RTSP Handoff mechanism by providing the flexibility in logically dividing a film into many sections dynamically. We have implemented the proposed RTSP Multiple Handoff mechanism in the Linux kernel 2.6.18 on LVS-CAD web cluster. In the performance evaluation, we have tested a variety of scenarios providing the streaming services with different proportions of short, medium, and long films on LVS-CAD and LVS web clusters. Experimental results demonstrate that the LVS-CAD web cluster using RTSP Multiple Handoff mechanism can achieve 24.43% better throughput than the one using RTSP handoff mechanism under LARD/RC dispatching policy, when the proportion of short, medium, and long films is 2:1:1. The throughput of LVS-CAD using RTSP Multiple Handoff mechanism also outperforms that of LVS without using our proposed RTSP Multiple Handoff mechanism under WLC and WRR dispatching policies by 36.34-107.52%. The experimental results also show that the web cluster using RTSP Multiple Handoff mechanism under LARD/RC dispatching policy can reduce 49.03-91.51% average response time compared with the one without using RTSP Multiple Handoff mechanism under WLC and WRR dispatching policies.[[note]]碩