Dynamic Job Replication for Balancing Fault Tolerance, Latency, and Economic Efficiency: Work in Progress

摘要

Recent research has demonstrated the benefits of replication of requests with canceling, which initiates multiple concurrent replicas of a request and uses the first successful result, immediately removing the remaining replicas of the completed request from the system. This paper suggests that the benefits of replication may come at the risk of an abrupt system transition to an undesirable highly congested equilibrium. To expose, evaluate, and ultimately manage these risk/benefit trade-offs, we generalize the replication strategy by: (a) accounting for the possible inefficiency of "remote" service, (b) allowing replication only when static routing fails to identify an idle "local" server, and (c) requiring one or more replicas of the same request to be completed to improve fault tolerance using a majority rule decision. Due to the intractability of the Markov performance model, our analysis is based on mean-field and fluid approximations. Future research should evaluate the accuracy of assertions based on these approximations, and ultimately develop practical solutions for optimization of various performance trade-offs in distributed systems with replication.