Runtime Adaptation of an Iterative Linear System Solution to Distributed Environments

Abstract

Distributed cluster environments are becoming popular platforms for high performance computing in lieu of single-vendor supercomputers. However, the reliability and sustainable performance of a cluster are difficult to ensure since the amount of available distributed resources may vary during the application execution. To increase robustness, an application needs to have self-adaptive features that are invoked at the runtime. For a class of computationally-intensive distributed scientific applications, iterative linear system solutions, we show a benefit of the adaptations that change the amount of local computations based on the runtime performance information. A few strategies for efficient exchange of such information are discussed and tested on two cluster architectures.