Impact of Frequency Scaling on One Sided Remote Memory Accesses

CPU Frequency scaling is a common approach used for achieving energy savings in parallel applications. A typical approach for achieving power savings is by reducing the frequency of a processor whenever the invested CPU cycles do not contribute to the progress of an application (e.g. polling for events). Many recent research efforts have been directed towards employing this approach within HPC applications. However, one of the important challenges that have remained unexplored is its potential impact on data transfer operations within communication-intensive applications.This paper presents a detailed analysis of the impact of frequency scaling on the performance and energy efficiency of one-sided point-to-point communication within PGAS models. As an an outcome of this study, we observe that the extent of this impact depends on a number of factors like the implementation approach of a data transfer operation and the number of explicit remote transfers initiated. Besides the characteristics of the software stack, hardware features like the microarchitectural design also impact the performance of such data movement while adopting voltage scaling techniques. These inferences are supplemented with empirical analysis of common approaches of implementing remote write (PUT) operations.The empirical results presented in this paper should act as a guide for future design of software solutions targeting energy efficiency in distributed systems.