A performance comparison of DSM, PVM, and MPI

Abstract

Abstract- I. INTRODUCTIONA number of computational problems cannot be solved in a reasonable time on currently available computers. Parallel computing provides a method to increase the performance of computationally intensive programs by distributing their execution across multiple processors. Traditionally this parallel processing was done on expensive purpose-built multiprocessor systems. An alternative which is becoming more popular is the use of multicomputer or cluster computer systems.A cluster computer system consists of a collection of computers connected by a network. Typically the computers are standard off-the-shelf personal computers. An example of such a system is a Beowulf cluster [1].Each node in the cluster has its own memory, accessible only to the processes running on that node. Programs running in parallel on different nodes require some way to communicate. The most common solution is the use of libraries which provide primitives to enable the parallel processes to communicate. Many of these libraries run in user space so no alteration of the operating system is required.To date, most of the libraries are based on the message passing paradigm. This requires a programmer to specify explicitly the data to be passed between processes and the time when the communications should take place. An alternative approach provides the abstraction of a single memory address space available to all processes running in the cluster. This approach is known as distributed shared memory (DSM). It frees a programmer from having to deal explicitly with data exchange, and therefore, provides an easier method of programming. The DSM library is responsible for memory consistency maintenance and ensures each process has access to the most recent version of data.Since cluster computers are fairly new, relatively little is known about the performance of DSM compared to message

Performance Evaluation of View-Oriented Parallel Programming on Cluster of Computers

“…However, if we compare the VOPP programs with the MPI programs, we find there are much more barriers in the VOPP programs [7]. Gauss calls 15998 barriers; IS calls 120 barriers; SOR calls 200 barriers; NN calls 2410 barriers; Water calls 72 barriers; TSP calls 3 barriers; Barnes-Hut calls 19 barriers; BT calls 2 barriers; CG calls 1186 barriers and MG calls 484 barriers.…”

“…Traditional DSM systems like TreadMarks [1] are far from efficient compared with Message Passing Interface [7]. The reason is that programs written in MPI can be finely tuned by reducing unnecessary message passing.…”

Performance Evaluation of View-Oriented Parallel Programming

“…A Distributed Shared Memory (DSM) system can provide application programmers the illusion of shared memory on top of message-passing distributed systems, which facilitates the task of parallel programming in distributed systems. However, programs using DSM are normally not as efficient as those using the Message Passing Interface (MPI) [12]. The reason is that message passing is part of the design of a MPI program and the programmer can finely tune the performance of the program by reducing the unnecessary message passing.…”