Iterators, Schedulers, and Distributed-memory Parallelism

In previous work, we demonstrated the advantages of encapsulating query evaluation algorithms as ‘iterators’ for sequential and parallel query evaluation. Unfortunately, those earlier models have a severe drawback with respect to resource allocation in distributed‐memory systems. Since threads may b...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Software, practice & experience Jg. 26; H. 4; S. 427 - 452
1. Verfasser: GRAEFE, GOETZ
Format: Journal Article
Sprache:Englisch
Veröffentlicht: New York John Wiley & Sons, Ltd 01.04.1996
Wiley
Schlagworte:
abstraction
Applied sciences
bottom-up activation
communication
Computer science; control theory; systems
Computer systems and distributed systems. User interface
database query processing
encapsulation
Exact sciences and technology
extensibility
Information systems. Data bases
iterators
Memory organisation. Data processing
parallelism
Software
software engineering
Encapsulation
Query
Resource allocation
Database
Parallelism
Data processing
Distributed system
Abstraction
Shared memory
Software engineering
ISSN:0038-0644, 1097-024X
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:In previous work, we demonstrated the advantages of encapsulating query evaluation algorithms as ‘iterators’ for sequential and parallel query evaluation. Unfortunately, those earlier models have a severe drawback with respect to resource allocation in distributed‐memory systems. Since threads may be initiated long before they actually perform useful work, thread placement decisions may be suboptimal. In this paper, we briefly review the iterator model and then extend it to support bottom‐up, just‐in‐time activation of appropriate query plan fragments as well as local and global synchronization and communication among sibling threads. Some of the algorithms described here may seem intricate; however, the intricacy is encapsulated entirely in the parallelism or ‘exchange’ iterator, thus freeing developers of data manipulation iterators to focus on the specific algorithms at hand, instead of on mechanisms for parallelism. Moreoever, code fragments extracted from an operational prototype demonstrate the ease with which our suggestions can be implemented. Run‐time overhead and performance of our techniques are identical to those of earlier, less extensible techniques for bottom‐up activation; the benefits of the presented techniques are their clean design, ion, extensibility, and the resulting ease of maintenance and enhancement.
Bibliographie:istex:C72EBE252447D1192CC36EFC23838AB79B81FD35
ArticleID:SPE20
ark:/67375/WNG-8PZ2LGCB-J
ObjectType-Article-2
SourceType-Scholarly Journals-1
ObjectType-Feature-1
content type line 23
ISSN:0038-0644
1097-024X
DOI:10.1002/(SICI)1097-024X(199604)26:4<427::AID-SPE20>3.0.CO;2-H