Modeling the slowdown of data-parallel applications in homogeneous and heterogeneous clusters of workstations

Data-parallel applications executing in multi-user clustered environments share resources with other applications. Since this sharing of resources dramatically affects the performance of individual applications, it is critical to estimate its effect, i.e., the application slowdown, in order to predict application behavior. The authors develop a new approach for predicting the slowdown imposed on data-parallel applications executing on homogeneous and heterogeneous clusters of workstations. The model synthesizes the slowdown on each machine used by an application into a contention measure-the aggregate slowdown factor-used to adjust the execution time of the application to account for the aggregate load. The model is parameterized by the work (or data) partitioning policy employed by the targeted application, the local slowdown (due to contention from other users) present in each node of the cluster and the relative weight (capacity) associated with each node in the cluster. This model provides a basis for predicting realistic execution times for distributed data-parallel applications in production clustered environments