Application-based energy efficient mobile and server computing
Energy efficiency is one of the major challenges faced by designers of computing systems. Mobile computing systems require energy efficiency to extend their battery lifetimes. Enterprise computing systems (servers) require energy efficiency to combat power and cooling problems in data centers. This dissertation examines how application-based improvements can provide energy efficiency for both mobile systems and servers, by addressing the following three questions. For mobile systems, I address: (1) under what conditions can a mobile system save energy by offloading parts of computation to a server, and (2) what criteria can be used for a mobile system to select a server for offloading. On the server side, I address (3) how an enterprise server can save energy by adapting its memory configuration based on application behavior. The first question is addressed by examining an improved implementation of content-based image retrieval (CBIR), and identifying the conditions for offloading to save energy. I show how application-based improvement of CBIR can influence offloading decisions. The second question is addressed by proposing a metric to select a server for offloading. The metric is given as the ratio of two factors: (a) the energy saved due to offloading and (b) the additional energy consumed for sending the data to a server and while waiting for the results. The third question is addressed by examining different factors in memory configurations such as capacity, frequency of operation, and replacement policies of row buffers. I present a case study showing the power-performance trade-offs in adapting memory capacity and frequency. I also present an algorithm that adapts the row buffer policy and saves 6% and 33% energy for two enterprise workloads.
Lu, Purdue University.
Off-Campus Purdue Users:
To access this dissertation, please log in to our