Title: Application based Evaluation of Distributed Shared Memory Versus Message Passing
Abstract: Clusters of Symmetrical Multiprocessors (SMPs) have recently become very popular as low cost, high performance computing solutions. While some programs can be automatically parallelized for use on a single SMP node, using multiple nodes in a cluster requires the programmer to rewrite the sequential code and employ explicit message passing. This paper explores an alternate approach, the use of a multithreaded Distributed Shared Memory (DSM) system, Strings. Though shared memory programs are easier to write, a DSM system may not perform as well as a message passing library. The performance of both approaches is evaluated using two applications from the field of medical computing. The first application is a program for deblurring images obtained from Magnetic Resonance Imaging. The other program is part of a system for radiation treatment planning. Each program was initially parallelized using a message passing approach. The programs were then rewritten to use a multithreaded approach over the DSM. Our results show that current implementations of the standard message passing libraries PVM and MPI are not able to effectively exploit multiple processors on an SMP node. Since Strings is multithreaded, it provides very good speed-up for both the programs in this environment. It is also seen that the DSM code is as good as the message passing version for one program, and nearly as good in the other program.
Publication Year: 2007
Publication Date: 2007-01-01
Language: en
Type: article
Access and Citation
Cited By Count: 1
AI Researcher Chatbot
Get quick answers to your questions about the article from our AI researcher chatbot