Analysis of a programmed backoff method for parallel processing on Ethernets
In many parallel processing applications, task times have relatively little variability. Accordingly, many nodes will complete a task at approximately the same time. If the application is run on an Ethernet, the near-simultaneity of the task completion times implies that when the tasks attempt to communicate with some central task manager, they will bump into each other. This in turn can cause a major slowdown in communication, as the Ethernet hardware generates unnecessarily long backoff times. The work here will analyze a solution to this problem.
Unable to display preview. Download preview PDF.
- 1.Vikram S. Adve and Mary K. Vernon, “The Influence of Random Delays on Parallel Execution Times,” Proceedings of the 1993 ACM Sigmetrics Conference on Measurement and Modeling of Computer Systems, May 1993, pp. 61–73.Google Scholar
- 2.S.G. Akl. “The Design and Analysis of Parallel Algorithms”, Prentice Hall, Inc, 1989.Google Scholar
- 3.G. Ciola, G. Ciaccio and L. Mancini. GAMMA Project: Genoa Active Message MAchine, Web page http://www.disi.unige.it/project/gamma.Google Scholar
- 4.Gregory Davies and Norman Matloff. “Network-Specific Performance Enhancements for PVM,” Proceedings of the Fourth IEEE International Symposium on High-Performance Distributed Computing, August 1995, pp205–210.Google Scholar
- 5.G. Gonnet. Handbook of Algorithms and Data Structures, Addison-Wesley, 1984.Google Scholar
- 6.Gilbert Held. Ethernet Networks (second edition), John Wiley, 1996.Google Scholar
- 7.G. Wilson. Practical Parallel Programming, MIT Press, 1995.Google Scholar