Adaptive scheduling of master/worker applications on distributed computational resources

One of the most popular ways to implement parallel operations on distributed-memory architectures is with a Master-Worker (MW) organization that concentrates control functions within a single master process, and delegates responsibility for computations to remote worker processes. While the MW approach is conceptually simple to put in practice, achieving consistent portable performance over a wide range of available distributed computing environments requires a variety of scheduling capabilities and techniques that allow application behavior to be tailored to fit specific environmental conditions. This dissertation addresses problems in simultaneously achieving MW application performance, portability, and ease of development.; Application performance is determined by the combination of specific application requirements and the available capacity of system resources to meet those needs. This dissertation presents a work-flow model of MW application performance that correctly accounts for both computation and communication-based performance constraints. Using this work-flow model, we have developed a resource selection algorithm for choosing appropriate hosts for master and worker processes that delivers performance up to levels allowed by application-specific constraints.; Application portability is essential for ensuring that a wide array of platforms are available to users as targets for individual MW applications. Instead of simply providing portability by reducing available capabilities to a least common denominator, we present an AppLeS Master Worker Application Template (AMWAT) approach to developing MW applications that maximizes program portability while also providing access to unique resource capabilities and specialized scheduling techniques.; We present a variety of basic and specialized scheduling techniques for MW applications, and then experimentally show how different techniques are appropriate in specific scheduling regimes. In particular, the effectiveness of different work distribution strategies are experimentally compared for a set of test applications and environmental conditions. We have incorporated each of the scheduling techniques into a portable reusable performance-oriented scheduler module.; Ease of MW application development is specifically addressed in the AMWAT approach. Rather than have every MW application implement many different common functions, we show how MW development can be simplified by separating functions provided by common components, such as the general scheduler module, from purely application-specific functionality.