An Emprical Evaluation of Memory Management Alternatives for Real-Time Java

Abstract

Memory management is a critical issue for correctness and performance of hard-real time systems. Java environments usually incorporate high-throughput garbage collection algorithms, but these algorithms can induce pause times in excess of 100 milliseconds. This is not acceptable for a real-time system. Two approaches for remedying this problem are being investigated. The pause times can be bounded with a real-time garbage collector; or a means to program around the collector entirely can be provided, as is done in the real-time specification for Java with its scoped memory interface. This paper presents the first side-by-side empirical evaluation of the impact of memory management regimes on realistic realtime applications. We present usability arguments as well as a detailed performance comparison of scoped memory and real-time garbage collection. Experience with medium sized systems suggests that while programming with scoped memory is error prone, it provides substantially better throughput. We have observed a throughput reduction of up to 37% and, in the worst-case, an 80% latency penalty for real-time garbage collection