Eventually-Consistent Federated Scheduling for Data Center Workloads

Abstract: Data center schedulers operate at unprecedented scales today to accommodate the growing demand for computing and storage power. The challenge that schedulers face is meeting the requirements of scheduling speeds despite the scale. To do so, most scheduler architectures use parallelism. However, these architectures consist of multiple parallel scheduling entities that can only utilize partial knowledge of the data center's state, as maintaining consistent global knowledge or state would involve considerable communication overhead. The disadvantage of scheduling without global knowledge is sub-optimal placements-tasks may be made to wait in queues even though there are resources available in zones outside the scope of the scheduling entity's state. This leads to unnecessary queuing overheads and lower resource utilization of the data center. In this paper, extend our previous work on Megha, a federated decentralized data center scheduling architecture that uses eventual consistency. The architecture utilizes both parallelism and an eventually-consistent global state in each of its scheduling entities to make fast decisions in a scalable manner. In our work, we compare Megha with 3 scheduling architectures: Sparrow, Eagle, and Pigeon, using simulation. We also evaluate Megha's prototype on a 123-node cluster and compare its performance with Pigeon's prototype using cluster traces. The results of our experiments show that Megha consistently reduces delays in job completion time when compared to other architectures.