PERFORMANCE ANALYSIS OF ON-THE-FLY GARBAGE COLLECTION.

Abstract

Parallel garbage collection systems consist of two processors, the mutator and the garbage collector, which operate on a common memory of list nodes. The garbage collector repeatedly executes a two-stage cycle that maintains a list of nodes (the free list) for use by the mutator, which is dedicated to executing a user's list processing program. All list nodes contain two extra bits for marking purposes and are either white, black, or gray. During the mark stage, the garbage collector blackens all nodes that are accessible to the mutator. During the scan stage, white nodes are placed on the free list and black nodes are whitened. The color gray is used during the mark stage to guarantee that all accessible nodes are marked. It is shown that parallel garbage collection systems exhibit three fundamental types of performance (stable, alternating, and critical), and it is determined when each of the three will occur. Moreover, it can be determined when this parallel scheme is more efficient than sequential garbage collection by obtaining analytic formulas for a variety of performance measures. These results are obtained by expressing the performance of parallel garbage collection systems in terms of conditional difference equations in which the state of the system after a garbage collection cycle is described by a pair of difference equations and a Boolean function. The value of the Boolean function at the beginning of a cycle selects which of the two difference equations should be applied next.