Data-parallel applications executing in multi-user clusteredenvironments share resources with other applications. Since this sharingof resources dramatically affects the performance of individualapplications, it is critical to estimate its effect, i.e., theapplication slowdown, in order to predict application behavior. Theauthors develop a new approach for predicting the slowdown imposed ondata-parallel applications executing on homogeneous and heterogeneousclusters of workstations. The model synthesizes the slowdown on eachmachine used by an application into a contention measure-the aggregateslowdown factor-used to adjust the execution time of the application toaccount for the aggregate load. The model is parameterized by the work(or data) partitioning policy employed by the targeted application, thelocal slowdown (due to contention from other users) present in each nodeof the cluster and the relative weight (capacity) associated with eachnode in the cluster. This model provides a basis for predictingrealistic execution times for distributed data-parallel applications inproduction clustered environments
展开▼