Shall I Use Heterogeneous Data Centers? - A Case Study on Video on Demand Systems

Recent research work and industrial experience show that heterogeneous data centers can be more cost-effective for some applications, especially those with intensive I/O operations. However, for a given application, it remains difficult to decide whether the application should be deployed on homogeneous nodes or heterogeneous nodes. In this paper, we perform a case study on finding optimized configurations for Video on Demand(VOD) systems. We examine two major modules in VOD systems, the streaming module and the transcoding module. In this study, we measure performance and power on different configurations by varying these factors: processor type, disk type, number of disks, operating systems and module mapping modes. We find that hosting VOD systems on heterogeneous nodes reduces no more than 35% power or cost per unit work done, compared with the server-only configuration. In Windows, multiplexing the two modules on servers achieves comparable power and cost reduction, but it is even worse than the non-multiplexing solution in Linux. The anti-intuitive and non-consistent results reflect the intrinsic complexity of finding an optimized configuration for a VOD system. Empirical data are necessary to support quantitative instead of qualitative analysis.