Distinguishing between prolific and non-prolific types for efficient memory management

In this paper we introduce the notion of prolific and non-prolific types, based on the number of instantiated objects of those types. We propose and empirically validate a new hypothesis, the prolific generational hypothesis, which states that the objects of prolific types have short lifetimes. We use this hypothesis to develop a new garbage collection scheme, which is similar in spirit to generational collection, but uses types rather than age to distinguish between different regions of the heap. It looks for garbage first among objects of prolific types. Our approach lets the compiler eliminate redundant write barriers by simply checking the type of objects involved in a pointer assignment. We describe techniques that enable this test to be done effectively at compile time, in spite of the complications due to polymorphism and dynamic class loading in Java. A preliminary implementation of this approach as a non-copying type-based collector in the Jalapeño Java Virtual Machine has shown very encouraging results. Compared with the corresponding generational collector, for the SPECjvm98 and SPECjbb2000 benchmarks, the number of dynamically executed write barriers is reduced by 18% to 74% (except for three programs, for which there is no reduction). The total garbage collection times are reduced by an average of 7.4% over all benchmark programs, with improvements of up to 15.2%. The overall performance improves modestly (by up to 2.8%) for all programs except for one (mtrt), for which there is a performance degradation of 2.2%.