Proposal : Analyzing Systems with Redundant Requests

Reducing response time is a primary focus of computer systems research. One key factor that influences a job’s response time in a multi-server system is dispatching: when a job arrives to the system, it must be sent immediately to one of the servers. This thesis focuses on a new dispatching policy: redundancy. Unlike traditional dispatching policies, which send only a single copy of each job, the idea of redundancy is to dispatch multiple copies of the same job and wait for the first copy to complete service. A great deal of empirical work has demonstrated that redundancy can provide substantial response time improvements. For example, using redundancy in MapReduce systems has been shown to reduce the response time of straggling tasks by 20 − 50%. However, despite the extensive empirical studies on redundancy, there has been very little theoretical work analyzing performance in redundancy systems. In this thesis, we propose to (1) quantify the benefits and costs of redundancy by providing the first analysis of systems with redundancy, and (2) design better redundancy systems to reduce response time. This proposal begins by reviewing our completed work, in which we develop exact analysis of response time in an idealized redundancy model. We use this analysis to explore initial messages on the benefits and costs of redundancy in small systems; we then analyze response time in systems with large numbers of servers. Our proposed future work consists of three major foci: developing and analyzing theoretical models of redundancy that capture the characteristics of real systems; investigating how smart scheduling policies can be used to improve performance in redundancy systems; and exploring applications for redundancy systems.