The networking research community has proposed several techniques for estimating the quality of network paths in terms of delay and capacity. However, hardly any have been studied in the context of really large applications deployed on the Internet. We have studied online gaming and have analyzed data from Halo 3 - a popular Xbox 360 game. Our dataset covers 5.6 million unique IP addresses that played 39.8 million game sessions over the Internet in 50 days. This application is extremely sensitive to network path quality, and Xbox360 consoles in each game session record latency and capacity measurements among themselves. We present several results from this analysis, including the global distribution of players, diurnal playing behavior, frequency of game play, temporal and geographical correlations in network performance. Based on the data, we propose three techniques to predict network path quality and hence reduce network overhead and improve scalability.