Date of Award

2001

Level of Access Assigned by Author

Open-Access Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computer Science

Advisor

James L. Fastook

Second Committee Member

Larry Latour

Third Committee Member

Thomas Wheeler

Abstract

Shared Virtual Memory (SVM) is an effort to provide a mechanism for a distributed system, such as a cluster, to execute shared memory parallel programs. Unfortunately, SVM has performance problems due to its underlying distributed architecture. Recent developments have increased performance of SVM by reducing communication. Unfortunately this performance gain was only possible by increasing programming complexity and by restricting the types of programs allowed to execute in the system. Validity resolution is the process of resolving the validity of a memory object such as a page. Current SVM systems use synchronous or deferred validity resolution techniques in which user processing is blocked during the validity resolution process. This is the case even when resolving validity of false shared variables. False-sharing occurs when two or more processes access unrelated variables stored within the same shared block of memory and at least one of the processes is writing. False sharing unnecessarily reduces overall performance of SVM systems?because user processing is blocked during validity resolution although no actual data dependencies exist. This thesis presents Asynchronous Validity Resolution (AVR), a new approach to SVM which reduces the performance losses associated with false sharing while maintaining the ease of programming found with regular shared memory parallel programming methodology. Asynchronous validity resolution allows concurrent user process execution and data validity resolution. AVR is evaluated by com-paring performance of an application suite using both an AVR sequentially con-sistent SVM system and a traditional sequentially consistent (SC) SVM system. The results show that AVR can increase performance over traditional sequentially consistent SVM for programs which exhibit false sharing. Although AVR outperforms regular SC by as much as 26%, performance of AVR is dependent on the number of false-sharing vs. true-sharing accesses, the number of pages in the program’s working set, the amount of user computation that completes per page request, and the internodal round-trip message time in the system. Overall, the results show that AVR could be an important member of the arsenal of tools available to parallel programmers.

Download

Share