在異質伺服器環境下利用乏晰/適應性類神經乏晰技術之負載平衡分配器

Abstract

隨著網際網路的流行，其流量日漸增加，尤其對於一個應用服務提供商(ASP)而言，它的流量更是龐大，當如果有很多個使用者同時連線時，在此時由於系統的容量有限，它無法完全快速滿足使用者的要求，此時就會發生擁塞。最好解決這種問題的方法在於使用大量的伺服器同時滿足使用者的要求，但由於成本的考量，應用服務提供商不可能無限制地增加伺服器的數量，因此在有限的伺服器數量下，依據伺服器的負載來分配新的使用者連線而使每個伺服器達到最大的系統效能則成為重要的問題。 在這幾年來，智慧型的控制技術如乏晰、類神經、適應性類神經乏晰理論皆已經大量廣泛地被應用在網路流量控制上，而許多的研究顯示這種作法大多都能夠改善傳統的方法。因此我們想到利用乏晰理論去解決負載平衡的問題。一般來說當我們知道現在伺服器的負載情形時，通常我們可以依此去判斷下一個時間該負載可能的情形，但是通常是不準確的，而且有不少誤差。在本論文中我們提出了乏晰演算法來解決這個問題。 我們的乏晰負載分配器是依據伺服器定時傳送過來的三個負載參數來做連線分配的決定，基於我們對於系統的瞭解，我們設計了乏晰系統中的乏晰規則及成員函數，這種利用乏晰理論並考慮到短時間及長時間的效能的方法，在我們的系統模擬中，能對於傳統的方法有不錯的改善。但這是基於我們對於系統的瞭解，這或許仍有所不足，因此我們利用適應性類神經乏晰理論去動態調整出最佳化的乏晰規則及成員函數，而利用適應性類神經乏晰理論這種技術，則又能改善一些系統效能。

The Internet traffic increases rapidly, especially the World Wide Web traffic. For a Application Service Provider(ASP), there are many users connect with it at the same time. Because the ability of server is limited, it cannot deal with all users' requests at the same time. Then the congestion is happened. The best solution to solve this problem is to increase the number of servers to deal with the users' requests and to duplicate the contents of the server. Because of the cost, the ASP cannot increase the number of servers unlimitedly. Under the condition of limited number of servers, it should dispatch new sessions of users to the server according to load balance of servers to maximize the system capacity and then avoid congestion. In recent years, The intelligent techniques such as fuzzy logic, neural network, ANFIS architecture, have been widely applied to deal with traffic control. Most research results show that the intelligent techniques can have better performance than conventional schemes. Now we have an idea to slove this problem by using fuzzy system. When we know all information of the current state, we may predict the load in the next state in some way. But it is not accuracy. In this thesis, we analyse the system deeply and propose the fuzzy algorithm. The dispatcher depends on the information which is sent periodically by the servers in the server farm to make decisions. This algorithm considers the short term and long term server characteristics to make decisions. This can have better performance than all other conventional algorithms in both request packet loss probability and overall system utilization. Based on the same model, we also propose an ANFIS technique to find the optimum fuzzy solution. The ANFIS technique can derive the best fuzzy rule and membership functions dynamically. And the performance of fuzzy algorithm in request packet loss probability can be further improved.