Java Garbage Collector Performance Measurements

Advances in compiler and virtual machine technologies support spread of programming environments with automated memory management, such as Java. These environments rely on garbage collection to reclaim memory that is no longer used. While in the domain of garbage collection research, every proposed garbage collector improvement is accompanied with detailed analysis of the collector overhead [Detlefs et al., 2004; Zhang and Seltzer, 2001; Blackburn, 2006], in the domain of performance modeling, the collector overhead is usually treated implicitly as a constant background factor [Xu et al., 2006; Becker et al., 2009; Kuperberg et al., 2008]. A potential reason for this treatment is lack of available information on the overhead of the garbage collector implementation. Although some implementations are open source, it is rather difficult to get such information from the code. For example, the garbage collector implementation of OpenJDK counts for more than 50000LOC. While this particular implementation is relatively well documented [Sun Microsystems, 2004], not all the information can be obtained from the documentation, and worse, the documentation contains misleading claims, such as that the JVM automatically adjusts total collector overhead to 1% of execution time [Sun Microsystems]. In order to provide a more precise collector overhead performance model, we have conducted many experiments to assess selected aspects of the garbage collector performance. Due to size constraints, the experiments are mostly reported online [Babka et al., 2009; Q-ImPrESS Consortium], and selected aspects are described in individual papers [Libič et al., 2009]. In this paper, we do not try to provide generalized claims, because the results of the experiments are bound to a particular virtual machine and collector. We do, however, compare two widely used collector implementations in the paper, and also attach notes on possible approaches towards modeling.