Statistical Analysis of the Workload of a Video Hosting Server

The amount of data hosted by Internet servers and data centers is increasing at a remarkable pace requiring more capable and more e cient servers. However, physical e ciency does not necessarily correlate with computational e ciency. In fact, independent studies reveal that Internet servers are mostly over provisioned even though additional servers are deployed each year. Understanding the characteristics of the workload of servers is an essential step to e ciently manage them. For example, from the workload statistics, it is possible to predict idle or underutilized states and to consolidate workload, so that the idle or underutilized servers can be switched o . In this paper, we systematically analyze the characteristics of video servers since they are responsible for producing the largest Internet tra c and provide an insight into the relationship between the statistics pertaining to workload, the size of videos, and service time. We shall show that from the distribution of the video sizes on host servers, it is possible to estimate the distribution of the workload size produced by clients and the distribution of the time required to process individual requests.