Reducing the Overhead of Compilation Delay

The execution model for mobile dynamically-linked object--oriented programs has evolved from fast interpretation to a mix of interpreted and dynamically compiled execution. The primary motivation for dynamic compilation is that compiled code executes significantly faster than interpreted code. However, since dynamic compilation is performed while the application is running, the biggest challenge in using dynamic compilation is to reduce its overhead so as not to mitigate the runtime improvement that it delivers. Techniques for reducing dynamic compilation overhead can be classified as (1) decreasing the amount of compilation performed, or (2) overlapping compilation with useful work. In this paper, we first evaluate the effectiveness of Lazy Compilation as a technique for decreasing the amount of compilation performed. In lazy compilation, individual methods are compiled on demand (when called), thus avoiding the load-time delay of compiling all methods when a new class/module is loaded. Our experimental results (obtained by executing the specJVM Java programs on the Jalapeno JVM) show that lazy compilation results in compilation of 57% to 63% fewer methods, and a reduction in compilation time of approximately 30%, when compared to load-time compilation. Next, we present Profile-driven Background Compilation as a new technique for overlapping compilation with execution. The motivation for background compilation is to use idle cycles in multiprocessor systems to overlap compilation with application execution. Profile information is used to prioritize methods as candidates for background compilation. Our results show that background compilation can deliver significant reductions (26% to 79%) in total time i.e., compilation plus execution time, compared to serial (non-background) compilation.