建構於雲端環境之多人線上遊戲動態資源分配機制

Abstract

大型多人線上遊戲是指數十萬的玩家同時上網進行遊戲。而運行大型多人線上遊戲時所消耗的CPU、記憶體以及網路頻寬等資源主要在客戶端玩家。我們將大型多人線上遊戲與雲端計算結合。在雲端計算環境下，我們利用虛擬伺服器來取代傳統實體伺服器。利用multi-server的遊戲架構，我們把虛擬遊戲世界切割成數個地圖區域，每個地圖區域由至少一個虛擬伺服器負責運行遊戲以及客戶端玩家間的訊息傳遞。我們根據每個虛擬伺服器的CPU、記憶體以及網路頻寬，利用類神經網路以及適應性類神經模糊系統來預測及決定該虛擬伺服器執行何種資源分配機制。這些資源分配機制包括：(1)可支援周圍虛擬伺服器；(2)解除被周圍虛擬伺服器支援的狀態或釋放支援本身的次要虛擬伺服器；(3)維持現有狀態；(4)須要請求周圍的虛擬伺服器支援；(5)在本身及周圍一個虛擬伺服器之間新增一個次要伺服器進行支援。根據我們的研究發現，適應性類神經模糊推論系統比起類神經網路有較低的均方根誤差，亦即有較好的學習效率。因此我們選擇適應性類神經模糊推論系統來實作上述五項資源分配機制。就資源分配機制而言，我們的方法比deep-level partitioning的存取時間快16.7%。

A massively multiplayer online game (MMOG) has hundreds of thousands of players who play in the game concurrently. The players consume a great deal of CPU, memory and network bandwidth resources in MMOGs. We combine MMOGs with cloud computing environments. We use virtual machine servers (VMSs) in cloud computing environments instead of traditional physical game servers. By using a multi-server architecture, we divide a game world into several zones, and each zone consists of at least a VMS to execute game processes and exchange game information among players in the zone. In addition, we design an artificial neural network (ANN) and also an adaptive neural fuzzy inference system (ANFIS) to predict the load of each zone and decide a resource allocation policy to be performed by the VMS. These policies include (1) this VMS is sufficient to support adjacent VMSs; (2) this VMS will release the resources which has been supported by adjacent VMSs or a secondary VMS which supports this VMS; (3) this VMS will remain in the current state; (4) this VMS requires adjacent VMSs to support it; (5) a secondary VMS will be created between this VMS and an adjacent VMS. Experimental results show that the mean square error of the ANFIS-based load prediction is lower than that of the ANN-based load prediction. Therefore, we incorporate the ANFIS prediction method along with the five resource allocation policies to the MMOG cloud. In terms of average access time, the proposed ANFIS-based resource allocation method is 16.7% better than the deep-level partitioning (DLP) method.