Impact of thread and frequency scaling on performance and energy in modern multicores: a measurement-based study

Abstract

Modern microprocessors integrate a growing number of components on a single chip, such as processor cores, graphics processors, on-chip interconnects, shared caches, memory controllers, and I/O interfaces. An ever-increasing complexity and the number of components present new challenges to software developers interested in finding operating points that strike an optimal balance between performance and energy consumed. In this paper we analyze the impact of thread scaling and frequency scaling on performance and energy in modern multicores. By exploiting recent additions to microprocessors that support energy estimation and power management, we measure execution times and energy consumed on an Intel Xeon 1240 v2 microprocessor when running the PARSEC benchmark suite. We conduct a number of experiments by varying the number of threads, 1 ≤ N ≤ 16, and processor clock frequency, 1.6 ≤ F ≤ 3.4 GHz. We find that the maximum performance is achieved when the number of threads matches or slightly exceeds the number of logical processors (8 ≤ N ≤ 12) and the clock frequency is at maximum (F = 3.4 GHz). The minimum energy is consumed when the processor clock frequency is in range 2.0 ≤ F ≤ 2.4 GHz. Finally, we find that the best performance at minimal energy is achieved when 8 ≤ N ≤ 12 and 2.8 ≤ F ≤ 3.1 GHz.