Communication Pattern Based Node Selection for Shared Networks

Selection of the most suitable nodes on a network to execute a parallel application requires matching the network status to the application requirements. We propose and validate a novel two step approach that exploits the knowledge of the communication structure of the application to address this problem. In the first step, a small set of candidate node groups are selected as potential sites of application execution, by analyzing the network status information and the communication patterns used by the application. The second step is based on the concept of a communication skeleton, which is a short running program that generates the dominant communication operations of the application it represents. The communication skeleton is executed on all candidate groups of nodes. The node group selected for application execution is the one that achieves the best performance on the communication skeleton. This approach leads to customized node selection and is particularly well suited to situations where available network information is of poor quality or expected communication performance cannot be modeled accurately. We motivate this approach, describe a prototype implementation, and present performance results for NAS Parallel Benchmarks executing on a shared network testbed.