Modelling of quality of experience for game on demand in cloud-based infrastructure

Abstract

Graphics Processing Unit (GPU) assisted cloud gaming is an emerging paradigm whereby computation-intensive video games can be played in real-time using thin clients. The keystrokes input by the user on the client device are relayed to the game server. The game server executes all the processes related to the game, including graphics rendering using the GPU, and the resulting frames are streamed as a video over a network to the client. The principal aim of this project is to create a model that can simulate a gaming cloud. To this end, a test bed was set up to represent a gaming cloud. The test bed consists of two computers connected directly to each other. The research gaming platform GamingAnywhere was used to transmit the game from the server computer to the client computer. The server computer also had the Network Emulator for Windows and Wireshark installed. The former was used to simulate various different network conditions as required, while the latter was used to capture and analyse the data being transferred between the client and the server. Nine factors were identified that influence the gaming experience. These can be broadly divided into three groups: those that can be controlled by the game service provider, such as the codec used to deliver the game; those that are controlled by the client, such as game genre and those that are defined by the network, such as available bandwidth. A test group of subjects were asked to play different games under different conditions and rate the experience using an adaptation of the Absolute Category Rating and the Degradation Category Rating defined by the ITU for evaluating video quality of multimedia applications. The data collected was used to develop two models. The first is a model based on an artificial neural network. It predicts the Quality of Experience as a Mean Opinion Score, based on the input factors. The other is a Naive Bayes classifier. It uses the input parameters to predict which level of opinion score the gaming experience falls into. Both models perform well when predicting the opinion score. The correlation between the predicted and the subjective opinion scores is 0.66 for the artificial neural network and 0.74 for the Naive Bayes classifier. The Naive Bayes classifier, therefore, gives slightly better results than the artificial neural network model.