On the efficacy, efficiency and emergent behavior of task replication in large distributed systems

摘要: Large distributed systems challenge traditional schedulers, as it is often hard to determine a priori how long each task will take complete on resource, information that input for such schedulers. Task replication has been applied in variety of scenarios way circumvent this problem. consists dispatching multiple replicas and using the result from first replica finish. Replication schedulers (i.e. employ replication) are able achieve good performance even absence tasks resources. They also smaller complexity than making them better suitable large systems. On other hand, waste cycles with not Moreover, extra consumption resources raises severe concerns about system-wide system multiple, competing This paper presents comprehensive study replication, comparing against information-based establishing their efficacy (the delivered application), efficiency amount wasted), emergent behavior schedulers). We introduce simple access control strategy can be implemented locally by resource greatly improves overall which compete