Incremental garbage collection in massive object stores

There are only a few garbage collection algorithms that have been designed to operate over massive object stores. These algorithms operate at two levels, locally via incremental collection of small partitions and globally via detection of cross partition garbage, including cyclic garbage. At each level there is a choice of collection mechanism. For example, the PMOS collector employs tracing at the local level and reference counting at the global level. Another approach implemented in the Thor object database uses tracing at both levels. The author presents two new algorithms that both employ reference counting at the local level. One algorithm uses reference counting at the higher level and the other uses tracing at the higher level. An evaluation strategy is presented to support comparisons between these four algorithms and preliminary experiments are outlined