Barrier Methods for Garbage Collection

Garbage collection algorithms have been enhanced in recent years with two methods: generation-based collection and Baker incremental copying collection. Generation-based collection requires special actions during certain store operations to implement the \write barrier." Incremental collection requires special actions on certain load operations to implement the \read barrier." This paper evaluates the performance of diierent implementations of the read and write barriers and reaches several important conclusions. First, the inlining of barrier checks results in surprisingly low overheads, both for the write barrier (2{6%) and the read barrier (< 20%). Contrary to previous belief, these results suggest that a Baker-style read barrier can be implemented eeciently without hardware support. Second, the use of operating system traps to implement garbage collection methods results in extremely high overheads because the cost of trap handling is so high. Since this large overhead is completely unnecessary, operating system memory protection traps should be reimplemented to be as fast as possible. Finally, the performance of these approaches on several machine architectures is compared to show that the results are generally applicable.