A high performance dynamic meta-scheduler with migration support for grid applications.

摘要

A central issue in Grid computing is the fact that resources (compute and communication fabrics) may vary over time, and thus scheduling decisions made by super (meta) schedulers may be pre-empted locally in response to a number of local conditions. Current superschedulers/scheduling schemes (such as AppLeS) develop essentially static schedules which are non-responsive to fabric dynamicism, and thus a task set scheduled by typical superschedulers is largely incapable of taking advantage of dynamic enhancements to the fabric or (worse) to pre-emption by an aggressive local scheduling decision. In the worst case, such pre-emption can result in premature termination of a task set. We present a novel superscheduling schema based on the widely-popular AppLeS and Globus Grid middleware which addresses the issue of dynamic rescheduling of Grid applications when an a priori schedule is challenged locally within the fabric. Specifically, we are concerned with the difficult problem of migrating dependent task sets robustly. We explore the use of non-blocking coordinated checkpointing and demonstrate that, subject to certain conditions, dependent task sets may be migrated robustly. Additionally, we show how AppLeS can be extended to function as a high-performance superscheduler i.e., as a superscheduler concerned with group performance rather than individual application performance. Finally, we demonstrate certain performance characteristics of the scheme using a Grid simulation tool developed at the University of Connecticut.