Modeling the effects of contention on application performance in multi-user environments
Silvia M. Figueira · 1997
This dissertation presents a model for estimating the effects of contention for resources on application behavior. This model provides a measure of application slowdown due to contention, which can be used for application scheduling in production heterogeneous systems. My thesis is that the effects of contention for resources on common distributed applications can be synthesized in a slowdown factor that is a function of a few, simple parameters. This dissertation develops models to estimate contention effects in various platforms. These models provide two types of factors: the slowdown factor, which synthesizes the effects of competing applications on computation and communication costs, and the aggregate slowdown factor, which synthesizes the effects of the global load in a cluster of workstations on a data-parallel task. These factors can be used by scheduling mechanisms in the evaluation of candidate schedules. The models developed are based on system characteristics and load behavior. In particular, the usage of the models depends on a clear characterization of both the competing and the target applications. This dissertation also presents a mapping strategy for high-performance applications executing on distributed heterogeneous systems. The strategy is based on a mapping model, the match-tree, which reflects the data movement and conversion costs of distributed algorithms. This strategy allows for both alternative implementations of individual tasks on different machines and inclusion of contention effects in performance estimates.