Using Kernel Instrumentation to Improve CPU Scaling Strategy

Traditionally, speed has been the key metric in evaluating computing systems. However, there are many environments in which energy eciency is also important. Such environments include mobile computing, which must budget a limited battery life, and high-performance computing, which often has a xed budget of incoming power or cooling. For this reason, it is often desirable to explore the tradeo between processing performance and power consumption. One technique to do this is to reduce the frequency and voltage of the CPU, which can result a quadratic boost in energy eciency. The key question here is when and to what extent to scale the CPU. In this paper, we explore a technique to determine when CPU frequency scaling would save energy. We use a variety of statistics to determine when the CPU is not the bottleneck of system performance, and we show that CPU scaling would be advantageous during such times. We extend previous work by using kernel instrumentation to detect IO-bound workloads suitable for scaling. We start with a system that examines metrics discussed in prior work. When we expand the system to include IO utilization, we nd that we miss 33.1 % fewer scaling opportunities and reduce the error rate by a factor of 4.3. Thus, we conclude that the addition of IO utilization signicantly improves our ability to predict when CPU scaling would be eective.