Power-performance analysis of JVM implementations

Java Virtual Machines (JVMs) work between the Java application programs and the operating systems (and their underlying hardware platforms) to provide the `write-once run-anywhere' property of the Java language. However, this property also implies that the runtime efficiency, in terms of both performance and power-consumption, can be affected by the implementations of JVM. In this paper, we present a case study of performance-power analysis of JVM implementations. We run SPECjvm2008 with OpenJDK and IBM I9 on an Atom-based netbook with Ubuntu operating system. Our observations are as follows: (1) the relative performance of OpenJDK ranges from 44 to 289% of J9, (2) the dynamic power consumption ranges from 2.8 to 7.2 Watts among benchmark programs. However, the power consumptions of two JVMs for the same workload are relatively similar. Therefore, the power-performance efficiency is mostly affected by the relative performance. (3) the effectiveness of multi-threading varies among benchmark programs as well as among JVMs. In general, running benchmark with a single-thread loses more in performance than in power consumption. Exceptions are compress, fft.small (OpenJDK only) and lu.small. (4) for most benchmark programs (except scimark), the power consumptions seem to be correlated to linear and square root of L2 reference and L2 miss rates, respectively.