Improving performance of Java applications using a coprocessor

Java technology is becoming a significant force stimulating the evolution of embedded systems. One of the problems In front of it is its poor performance. An important component of this performance problem is the overhead time spent during dynamic compilation. In this paper, we demonstrate that a dual-processor heterogeneous architecture that employs both a conventional CPU and a Java coprocessor dedicated to compilation can be very effective in hiding the time spent in dynamic compilation. We propose two different profile-driven algorithms that make use of such a coprocessor based system, and evaluate them using a set of Java benchmarks. Our results clearly indicate that it is possible to improve overall performance by intelligently scheduling dynamic compilations on the coprocessor.