Infant mortality and generational garbage collection

Abstract

Generation-based garbage collection has been advocated by appealing to the intuitive but vague notion that "young objects are more likely to die than old objects". The intuition is, that if a generation-based garbage collection scheme focuses its effort on scanning recently created objects, then its scanning efforts will pay off more in the form of more recovered garbage, than if it scanned older objects. In this note, we show a counterexample of a system in which "infant mortality" is as high as you please, but for which generational garbage collection is ineffective for improving the average mark/cons ratio. Other benefits, such as better locality and a smaller number of large delays, may still make generational garbage collection attractive for such a system, however.