Fortune teller: improving garbage collection performance in server environment using live objects prediction

Currently, the most adopted criterion to invoke garbage collection is heap space exhaustion. In other words, garbage collection is invoked when the heap space (either the entire space or generational space) runs out. A possible alternative but much more difficult approach is to invoke garbage collection when good garbage collection efficiencies can be obtained. To do so, we need to know number of garbage objects that can be collected before the collection actually takes place.In this short abstract, we introduce a predictive approach that can provide an accurate estimation of the number of dead objects at any specific point of execution. The proposed estimation model relies on the information obtained from partial reference counting. Our plan is to use this information as a criterion to invoke garbage collection. We have conducted a preliminary study to determine the feasibility of this idea and found that the model is sufficiently accurate in three SPECjvm98 benchmark applications.