skip to main content
Department of Computer Science University of Colorado Boulder
cu: home | engineering | mycuinfo | about | cu a-z | search cu | contact cu cs: about | calendar | directory | catalog | schedules | mobile | contact cs
home · events · thesis defenses · 1997-1998 · 
 

Thesis Defense - Aggarwal

 
1/9/1998
10:00am-12:00pm
ECOT 831

Analysis of the Hemingway Multiple-Writer Implementation of Release Consistency for Software Distributed Shared Memory Systems
Computer Science PhD Candidate

Software distributed shared memory systems benefit from implementations of weak memory consistency models because these allow for fewer messages to keep global memory coherent. A Release Consistent view of global shared memory offers potential for large degrees of concurrency, by reducing memory coherence to specific synchronization times. Release Consistency, in turn, can be implemented in many different ways.

We investigate the performance of schemes which allow for multiple writers to concurrently modify shared memory segments. These schemes alleviate false-sharing problems which plague single-writer schemes. However, although multiple-writer schemes may reduce the number of coherency operations, they are likely to require longer-latency synchronization operations. We investigate different techniques to exploit the reduction of false-sharing offered by multiple-writer protocols, while preserving the low-latency synchronization operations characteristic of single-writer schemes. We do this study in the context of different network bandwidths and sharing granularities.

Our results indicate that single-writer protocols provide good performance in the presence of efficient support for fine-grained sharing. Multiple-writer protocols, however, tolerate false-sharing, making the design of systems with coarse-grained sharing, effective. Further, we found that variations in network bandwidths do not affect the performance of write-through protocols significantly: latency of memory transfers between processor nodes is the important factor. Finally, our results suggest that clustering together processor nodes to form small-scale shared memory multiprocessors reduces memory transfer times and leads to large improvements in performance.

Committee: Dirk Grunwald, Associate Professor (Chair)
Daniel Scales, Compaq Computer Corporation
Gary Nutt, Professor
Andrew Pleszkun, Department of Electrical and Computer Engineering
Benjamin Zorn, Associate Professor

 
See also:
Department of Computer Science
College of Engineering and Applied Science
University of Colorado Boulder
Boulder, CO 80309-0430 USA
Questions/Comments?
Send email to

Engineering Center Office Tower
ECOT 717
+1-303-492-7514
FAX +1-303-492-2844
XHTML 1.0/CSS2 ©2012 Regents of the University of Colorado
Privacy · Legal · Trademarks
May 5, 2012 (13:40)
 
.