this paper, we first address fundamental issues in building useful performance-tuning tools and then describe our experience with the AIMS toolkit for tuning parallel and distributed programs on a variety of platforms. AIMS supports source-code instrumentation, run-time monitoring, graphical execution profiles, performance indices and automated modeling techniques as ways to expose performance problems of programs. Using several examples representing a broad range of scientific applications, we illustrate AIMS' effectiveness in exposing performance problems in parallel and distributed programs

Using a Receiver Makes it Right (RMR) data conversion technique in PVM significantly improves the message-passing performance in heterogeneous environments. The improvements are due to 3 factors: 1). RMR reduces the need for conversions in a heterogeneous environment; 2). At most each message is converted only once compared to twice for XDR used in current public version of PVM; 3). Our conversion routines are streamlined 1 and are several times faster than the XDR routines. The drawback to RMR is the potential need for a large number of conversion routines. We demostrate that only a small number of routines are required because many vendors use the IEEE standard for data representation. Given this fact, RMR may emerge as a promising technique in distributed computing. 1 Introduction PVM (Parallel Virtual Machine) is a software system that enables a heterogeneous collection of parallel and serial computers to be programmed as a single machine and utilized as a unified general and fl...

The limited communication bandwidth and high startup latencies of clustered workstations restrict their use to problems with sparse communication patterns or good concurrency between calculation and communication. First we describe our modifications to the popular PVM[5] message passing library, and on performance improvements using the PVM package on an FDDI-ring. Applications developed with a parallel communications architecture in mind perform poorly when ported to a message passing library running on workstations with sequential communication. In the second part, we present a dynamic loop scheduling algorithm for the data parallel programming model which optimizes the network usage on such clusters. As a proof of concept we have implemented a basic matrix multiplication and find a significant increase in parallel efficiency. 1 Introduction The promise to collect computer cycles from idle workstations has raised the interest in software packages which offer a message passing progra...

Processor Scheduling In A Distributed-Memory Computing Environment By Stephen W. Turner In recent years, the development of large-scale distributed-memory computers has given the user community unprecedented levels of computing power. In order to effectively use the available computing power, processor scheduling algorithms have been developed that allow many users to share distributed computing resources while obtaining the best possible job turnaround time. However, not all existing scheduling techniques take full advantage of available computing power. For example, in hypercubes, a cluster must normally be allocated as an entire subcube, which can result in high internal fragmentation, as well as poor job performance. Although the distributed workstation environment has recently become popular as a choice for a distributed-memory parallel computer, the problem of scheduling specifically for parallel job execution has not been well studied in this environment. In this thesis, we pr...