突波式流量之網頁伺服器負載平衡架構的研究

A Study of Web Server Load Balancing Architecture for Burst Mode Traffic

陳建伯(Jian-Bo Chen)

指導教授 ： 包蒼龍

大同大學/電機資訊學院/資訊工程學系所/博士(2008年)

摘要

伺服器負載平衡架構是經常用來解決熱門網站負載分攤的一種方法，而負載平衡系統有許多不同的解決方案。在本論文中，我們採用一種有彈性的註冊協定，使得負載平衡系統能夠很容易增加後端伺服器來分擔負載。除此之外，在這註冊協定的訊息中，同時也將伺服器的即時負載量回報給負載平衡系統，利用這些資訊，便可採用任一可行的負載平衡演算法來分配使用者的需求。
負載平衡演算法的目的是為了改善熱門網站的效能。大部分的負載平衡架構都僅能適用在同質性的後端伺服器上。如果後端伺服器的硬體規格不同時，負載平衡系統必須要有一種策略來將用戶端的連線需求，公平地分配到每一台後端伺服器上。我們推導一個異質性後端伺服器運算能力的度量機制。這些運算能力可以根據在特定丟棄率的條件下，後端伺服器所能提供的最大連線數來決定。然而，根據實驗的結果可知，這種定義方式無法保證每個用戶端均有公平的連線回應時間。因此，在考慮定義運算能力時，我們不僅是要考慮到丟棄率，同時也要考慮到回應時間。採用這樣的定義方式，對於所有用戶端連線的平均回應時間幾乎都是相同的。
除了定義運算能力之外，當網站並不是經常性的有突波連線需求時，我們可以採用剩餘能力負載平衡演算法來降低負載平衡系統中的硬體成本。我們提出一種隨選服務(service-on-demand)伺服器的概念，也就是在有突波連線需求時，將其他的伺服器像是DNS伺服器或是MAIL伺服器等，加入到負載平衡系統中來分擔流量。舉例來說，學校的選課系統或是訂票系統等，一年之中會有突波需求的次數僅有少數幾次而已。這種剩餘能力演算法可以利用DNS伺服器或是MAIL伺服器的剩餘能力，當突波連線需求來臨之前，加入到負載平衡系統中來分擔流量。模擬的結果發現，當同時使用DNS伺服器以及MAIL伺服器的剩餘能力時，所得到的結果，會比使用另一台專屬後端伺服器的效果來的要好。
由於每一台後端伺服器的剩餘運算能力均不相同，因此需要一種智慧型的機制來做決策。我們提出一種模糊決策的演算法，來將使用者的連線需求派遣給最適合的後端伺服器處理。每一台後端伺服器的CPU的閒置百分率、可用的記憶體空間以及可用的連線數，當成是模糊決策演算法的三個輸入參數。根據這些參數來計算最後的明確值以決定最適合的後端伺服器了。模擬的結果發現，當採用模糊決策負載平衡演算法時，可以比其他的演算法，得到更高的效能。

關鍵字

負載平衡 ； 剩餘能力 ； 模糊決策 ； 隨選服務

並列摘要

The server load balancing architecture is the most efficient way to solve the heavy loading problem of popular server. There are different solutions in implementing the load balancing system. In this dissertation, we adopt a flexible registration protocol that can easily add a new backend server to the load balancing system to share the load. In addition, the registration protocol also reports the real time backend server loading status to the load balancing system. So we can use these information to distribute client requests by any available load balancing algorithms.
The purpose of load balancing algorithm is to improve the load sharing performance of the popular web server. Most of the load balancing architectures are based on supporting homogeneous backend servers. If the hardware specifications of backend servers in the system are different, the load balancing system must have a strategy to fairly dispatch the load to the backend servers. We derive a capacity measurement for heterogeneous backend servers. From the experimental results, the maximum number of connection with certain drop rate can be used as the capacity, but it can not provide fair response time for each client requests. Thus, we must consider the capacity not only depending on drop rate but also on the response time. Using this measurement, the average response time for all client requests will nearly be the same.
In addition to the definition of capacity, we also use the remaining capacity algorithm to reduce the hardware cost when the web site does not have frequent burst requests. We propose the concept of service-on-demand servers, which can bring other servers such as DNS servers or MAIL servers to join the load balancing system during the burst traffic period. For example, the course registration system or ticket reservation systems have burst requests only several times a year. The proposed algorithm can use the remaining capacity of DNS or MAIL server to share the load in the burst request period. The simulation results show that using the remaining capacity of both DNS server and MAIL server can achieve higher performance compared to using another dedicated backend server.
Due to the different remaining capacities that these servers have, we need an intelligent mechanism to make the load distribution decision. We propose an algorithm using the fuzzy decision algorithm to dispatch the client requests to the appropriate backend server. The CPU idle percentage, the available memory percentage, and the available connection percentage for each backend servers are the input parameters of our fuzzy decision algorithm. The most appropriate backend server can thus be determined. The simulation results show that the fuzzy decision algorithm can achieve higher performance than other load balancing algorithms.

並列關鍵字

load balance ； remaining capacity ； fuzzy decision ； service-on-demand

參考文獻

[1] Eunmi Choi, Yoojin Lim, and Dugki Min, “Performance Comparison of Various Web Cluster Architectures,” Lecture Notes in Computer Science, vol. 3398, pp. 617-624, May 2005.

[2] Dan Mosedale, William Foss, and Robert Martin McCool, “Lessons Learned Administering Netscape’s Internet Site,” IEEE Trans. Internet Computing, vol. 1, no. 2, pp. 28-35, March/April 1997.

[3] Eric Dean Katz, Michelle Butler, and Robert McGrath, “A Scalable HTTP Server: The NCSA Prototype,” Trans. Computer Networks and ISDN Systems, vol. 27, no.2, pp. 155-164, 1994.

[5] Jeffery S. Chase, “Server Switching: Yesterday and Tomorrow,” Proc. Second IEEE Workshop on Internet Applications, pp. 114-123, San Jose, CA, 2001.

[14] Jiani Guo and Laxmi N. Bhuyan, “Load Balancing in a Cluster-Based Web Server for Multimedia Applications,” IEEE Trans. Parallel and Distributed Systems, vol. 17, no. 11, pp. 1321-1334, November 2006.