Diffusion tree restructuring for indirect reference counting

Abstract

A new variant algorithm for distributed acyclic garbage detection is presented for use in hybrid garbage collectors. The existing fault-tolerance of Piquer's Indirect Reference Counting (IRC) is qualitatively improved by this new approach. The key insight that underpins this work is the observation that the parent of a node in the IRC diffusion tree need not remain constant. The new variant exploits standard mechanisms for implementing diffusion trees and remote references, using four simple low-cost techniques to dynamically restructure the trees to reduce their depth. This variant reduces third-party dependencies, which make standard IRC vulnerable to process failure, while retaining tolerance of message reordering and without incurring substantial overheads. The paper carefully motivates the algorithm, presents the full technical basis for its development, provides a clear explanation of implementation details and includes an initial discussion of performance issues.

Birrell's distributed reference listing revisited

“…Many algorithms have been designed to combat the problem inherent in naive distributed reference counting, including Weighted Reference Counting [Bevan 1987;Watson and Watson 1987], Indirect Reference Counting [Piquer 1991], and those of Lermen and Maurer [1986], Birrell et al [1993], Moreau and Duprat [2001], and Dickman [2000]. We discuss some of these in the related work section, but we focus here on Birrell's algorithm, which is used in the distributed garbage collector for Java RMI.…”

“…Moreau's algorithm [Moreau 2001b;Moreau and Duprat 2001] and Dickman's [2000] algorithm show how zombies can be removed by reorganizing the diffusion tree by reparenting child nodes, either with the root node (Moreau) or arbitrarily (Dickman). Indeed, IRC can be seen as a special case of Moreau's algorithm.…”