Modeling Real-time Garbage Collection Cost

Given the major advantages of productivity and safety, the use of garbage collection (GC) in real-time systems has gained increasing attention. Guaranteeing garbage collection activities' worst-case execution time (WCET) is necessary for a real-time system to perform scheduling and schedulability analysis. This paper describes a detailed GC cost model for incremental mark-and-sweep GC exemplified by a modified Boehm-Demers-Weiser (BDW) collector. The GC cost model computes the WCET for garbage collection in terms of (1) the performance of collector operations and (2) the garbage collection load offered by a real-time task. Separating these aspects is a step along the road to an engineering approach to garbage collection in real-time systems, allowing prediction of system behavior from knowledge of component behavior and environmental specifications. The model incorporates the cost of write barriers that are needed to support incremental GC. To be useful for real-time systems, a model's predicted WCET must be no less than the actual WCET. However, unreasonably large overestimates are also problematic as they may lead to the incorrect conclusion that the application cannot be feasibly scheduled. A pessimism metric assesses the degree to which the model's predictions exceed the measured worst-case cost in particular instances.