GC-Consistent Cuts of Databases

We introduce a new method for concurrent mark-and-sweep garbage collection in objectoriented databases. For this purpose, we define a cut of a database to be a collection containing one or more copies of every page in the database; the copies may have been made at different times during the operation of the database. We define a class of cuts called GC-consistent cuts, and prove formally that a garbage collector can correctly determine which objects to delete by examining a GC-consistent cut of a database instead of the database itself. We show that GC-consistent cuts can synchronize the concurrent collector with the mutator, i.e. perform the task usually assigned to a write barrier: while a database is in operation, a GC-consistent cut of it can be built in an efficient and inobtrusive way, and, while still under construction, can be used by a garbage collector. We investigate other fundamental properties of GC-consistent cuts. We compare their consistency properties with those of causal cuts of distributed systems. We show that although the reachability of objects in a GC-consistent cut is inherited from the underlying database, many other interesting properties of the cut are unrelated to those of the database; this weak consistency is related to the low cost of building GC-consistent cuts. Key-words: concurrent garbage collection, object-oriented database, causal cut, GC-consistent cut, write barrier, distributed system.