Lifetime-based garbage collection algorithm is considered as the most successful among all garbage collection techniques. In this technique, the address space is divided into several partitions according to the lifetimes of objects. In this case, garbage collection cost depends on the number of partitions. Assuming the distribution of lifetimes of objects as a parameterized exponential probability density function, studied are the effects of the number of partitions an variable partition size on garbage collection cost.
Garbage collection cost is computed from the probability of an object to become a garbage at each level and the number of collections invoked at each level to finish a program. Experimental results shows that in almost all cases, lifetime-based garbage collector with one ephemeral level yields the minimum cost. However, when the objects are relatively stable, Baker's garbage collector, which is equivalent to zero ephemeral level, is better. By increasing the size of the first ephemeral level into 3 or 4 times bigger than others, garbage collection cost is reduced substantially.
Note that garbage collection cost is defined to be proportional only to the number of active cells. However, when using very large virtual memory systems, the definition of garbage collection cost should be modified to consider not only the number of active cells but also the overhead of page faults caused by uncollected garbages. A way to consider this overhead into the lifetime-based garbage collection cost is proposed.
