Application Development and Execution in a VirtualComputing

As the size and complexity of high performance computing applications continue to increase, their computing, storage and connectivity requirements become more and more di cult to accommodate on any single computing platform. The Virtual Computing Environment (VCE) combines heterogeneous computing, storage and network resources into a exible system for the execution of large scale applications. Current systems for managing distributed applications employ static allocation policies which are di cult to scale to large networks made up of geographically dispersed resources connected by high speed network technology (e.g. ATM). In this paper we present an application description framework which is used to develop large scale applications in the VCE. We also discuss two methods for scheduling tasks in the VCE: one emphasizes redundant task scheduling while the other emphasizes a more conventional use of load measurement techniques to select less loaded computers to run application tasks. Several example applications are included and used to demonstrate the performance of a Virtual Computing Environment which employs an IBM SP2 supercomputer, a cluster of DEC Alpha workstations, and a cluster of Sun and Silicon Graphics workstations interconnected by an ATM network. 2