Abstract: "We describe a concurrent, atomic garbage collection algorithm for transaction-based languages, a class of languages intended to support reliable distributed systems. A garbage collection algorithm for reliable distributed systems must be atomic: a crash during a garbage collection sh...

Ausführliche Beschreibung

Bibliographische Detailangaben

1. Verfasser:	Detlefs, David L. (VerfasserIn)
Format:	Buch
Sprache:	English
Veröffentlicht:	Pittsburgh, Pa. Department of Computer Science, Carnegie Mellon Univ. 1990
Schriftenreihe:	School of Computer Science <Pittsburgh, Pa.>: CMU-CS 1990,177
Schlagworte:	Datenverarbeitung Electronic data processing > Distributed processing Garbage collection (Computer science) Multiprocessors Speicherbereinigung Algorithmus Parallelverarbeitung Hochschulschrift
Zusammenfassung:	Abstract: "We describe a concurrent, atomic garbage collection algorithm for transaction-based languages, a class of languages intended to support reliable distributed systems. A garbage collection algorithm for reliable distributed systems must be atomic: a crash during a garbage collection should result in no loss of data. A concurrent collector allows programs to continue operating as their heaps are collected, minimizing delays imposed by collection. A concurrent collector can be used in reliable distributed systems that must service interactive requests We introduce concurrency into an atomic collector by breaking collection into a series of garbage collection segments, which are similar to transactions, but satisfy weaker properties than are required by transaction semantics. These weaker properties allow better performance. Thus, our collection algorithm enhances the reliability of programs written in transaction-based languages, without necessarily imposing severe performance penalties. We also describe a concurrent "mostly-copying" collection algorithm for C++, and present measurements of the performance of an implementation of this algorithm We survey garbage collection techniques applicable to "systems" languages such as C++, and show how these techniques may be combined with concurrent collection. Performance measurements of the resulting collector (on both a uniprocessor and a multiprocessor) indicate that interruptions caused by concurrent collection are quite short, and that adequate mutator performance is sustained during collection. Finally, we prove the correctness of most aspects of the concurrent, atomic collection algorithm. The proof hinges on showing that the routines of the algorithm preserve a novel property called partial idempotence, which ensures that partial effects of routines interrupted by crashes can be undone or completed during recovery.
Beschreibung:	Zugl.: Pittsburgh, Pa., Univ., Diss., 1990
Beschreibung:	XI, 218 S. graph. Darst.