Java garbage collector selection using object oriented software metrics

Java programming language introduces garbage collection to manage memory automatically. In this thesis, it is proposed that there should be a relation between garbage collector performance and program properties. It is also proposed that this relation can be exploited and good performing garbage collector for a program can be predicted by analyzing program. Predictor function can be constructed by using a machine learning system which creates a decision tree. Training data is needed to feed such system. Programs can be tested with different garbage collectors. After that program property data can be merged to garbage collector performance data to produce training data. The DaCapo Benchmark Suite is chosen for work. Benchmark programs are analyzed both statically and dynamically. Metrics including Chidamber and Kemerer metrics are extracted as static properties and also object allocation metrics extracted as dynamic properties through a single program execution. Program properties mentioned constructed training data with garbage collector results. Weka machine learning tool is used to build a decision tree. J48, which is an implementation of C4.5 algorithm is used to build a decision tree. Decision tree's success rate is measured as 76.85%. With this work, relation between program properties and garbage collection performance is shown. Using this relation a predictor system is proposed. This system had been able to classify instances correctly. In future research, it is possible to obtain better results with extended set of sample programs and program properties.