Hierarchical distributed reference counting

Abstract

Massively distributed computing is a challenging problem for garbage collection algorithm designers as it raises the issue of scalability. The high number of hosts involved in a computation can require large tables for reference listing, whereas the lack of information sharing between hosts in a same locality can entail redundant GC traffic. In this paper, we argue that a conceptual hierarchical organisation of massive distributed computations can solve this problem. By conceptual hierarchical organisation, we mean that processors are still able to communicate in a peer to peer manner using their usual communication mechanism, but GC messages will be routed as if processors were organised in hierarchy. We present an extension of a distributed reference counting algorithm that uses such a hierarchical organisation. It allows us to bound table sizes by the number of hosts in a domain, and it allows us to share GC information between hosts in a same locality in order to reduce cross-network GC traffic.