Flexible load balancing software for parallel applications in a time-sharing environment

Networks of workstations become more and more appropriate for parallel applications, as modern network technology enables high quality communication between powerful workstations. In this perspective, load balancing software must be extremely exible as the set of available nodes for a particular distributed memory application may change at run time. XENOOPS is an advanced environment for parallel software. The XENOOPS load balancing framework meets the requirements of the regular users of a time sharing distributed system, and of the HPCN users who prefer to exploit multiple processors as if these processors have been reserved for one particular application. On the one hand, a parallel application can dynamically obtain and eeectively exploit workstations as these turn idle. On the other hand, workstation owners transparently claim their device when accessing it on the console. In both cases, the parallel application transparently reorganises itself while maintaining a balanced work load distribution. Consequently, dynamic load balancing also becomes important for regular applications, which would not require dynamic reallocation in a space-sharing system. This paper treats the elements of XENOOPS that realise generic load balancing support. A key element is object migration combined with granularity control. We discuss the subsystem that integrates XENOOPS applications in a time sharing environment. We illustrate the resulting possibilities with a visualisation of dynamic load balancing behaviour in a PDE solver with adaptive grid reenement.