Evaluating partition selection policies using the PMOS garbage collector

Abstract

Implementors of garbage collection algorithms are, in general, faced with determining a number of policy decisions such as when to collect, how to collect space or how to interact with the running application. With main memory collectors such decisions are usually influenced by results and experiences from a wealth of previous work. However, with partitioned collection of persistent objects stores, the implementor has a much smaller base to draw from. This is due not just to the small number of existing incremental object store collector implementations but also because the tradeoffs are significantly different from main-memory collection. This paper reports on investigations, using the PMOS collector, into policy choices determining which partition should be chosen for collection. A previous study on partition selection postulated that a flexible selection policy can significantly reduce I/O and increase the amount of space reclaimed. That study was based on simulations on a synthetic object database. Here we test these conclusions by repeating the experiments using real applications, a partitioned store collector and an implemented persistent object store.