INCREMENTAL MULTI-THREADED GARBAGE COLLECTION ON VIRTUALLY SHARED-MEMORY ARCHITECTURES

This paper describes a multi-threaded and incremental garbage collector operating on shared memory architectures. The technique was developed for parallel implementations of the language LCS, a high level parallel programming language. An incremental, page trap based, collection algorithm operates locally on each of the processors. Processors alternatively plays the role of mutator and collector. The processors cooperate for collection and mutation; idling processors perform part of the collection task for the others until they acquire some work. The progress of collectors versus allocations is controlled by a scan credit mechanism that guarantees a responsive execution of the application. There is no static partitioning of the storage among the processors; pages are dynamically allocated to any processor, for specific purposes. Two implementations are discussed: the first is suitable for operation on a shared memory architecture; the second provides garbage collection services as added functionality to a distributed shared virtual memory service.