Phurti: Application and Network-Aware Flow Scheduling for Multi-Tenant MapReduce Clusters

Traffic for a typical MapReduce job in a data center consists of multiple network flows. Traditionally, network resources have been allocated to optimize network-level metrics such as flow completion time or throughput. Some recent schemes propose using application-aware scheduling which can shorten the average job completion time. However, most of them treat the core network as a black box with sufficient capacity. Even if only one network link in the core network becomes a bottleneck, it can hurt application performance. We design and implement a centralized flow-scheduling framework called Phurti with the goal of improving the completion time for jobs in a cluster shared among multiple Hadoop jobs (multi-tenant). Phurti communicates both with the Hadoop framework to retrieve job-level network traffic information and the OpenFlow-based switches to learn about the network topology. Phurti implements a novel heuristic called Smallest Maximum Sequential-traffic First (SMSF) that uses collected application and network information to perform traffic scheduling for MapReduce jobs. Our evaluation with real Hadoop workloads shows that compared to application and network-agnostic scheduling strategies, Phurti improves job completion time for 95% of the jobs, decreases average job completion time by 20%, tail job completion time by 13% and scales well with the cluster size and number of jobs.