A New Non-Blocking Counter-Based Coordinated Checkpointing Algorithm as a Migration Tool in a High Performance Dynamic Grid Scheduler

摘要

Scheduling computations and resource management for Grid applications is challenging as they are distributed, heterogeneous in nature, owned by different individuals or organizations with their own policies, have different access and cost models, and have dynamically varying loads and availability. High Performance Schedulers promote the performance of individual applications by optimizing performance measures such as minimal execution time, resolution, or speed-up. A central focus of this paper involves addressing the problem of migrating task sets on to (and off of) the compute fabric in response to changes in the fabric makeup (such as dynamic resource availability) or to global priority changes within the fabric and/or task sets. This issue is complicated if a task set to be migrated consists of dependent (messaging) subtasks. We present a new non-blocking counter-based coordinated Checkpointing algorithm for a dependent/ messaging task set. This algorithm obtains a global consistent checkpoint that can be used to migrate these tasks upon any performance degradation to other available computing resources and resumes their execution using the last checkpoint image. Also there is no restriction on the frequency of inter-task messages