Deadline-Aware Reservation-Based Scheduling

Abstract

The ever-growing need to improve return-on-investment (ROI) for cluster infrastructure that processes data which is being continuously generated at a higher rate than ever before introduces new challenges for big-data processing frameworks. Highly complex mixed workload arriving at modern clusters along with a growing number of time-sensitive critical production jobs necessitates cluster management systems to evolve. Most big-data systems are not only required to guarantee that production jobs will complete before their deadline, but also minimize the latency for best-effort jobs to increase ROI.

This research presents DARSS, a deadline-aware reservation-based scheduling system. DARSS addresses the above-stated problem by using a reservation-based approach to scheduling that supports temporal requirements of production jobs while keeping the latency for best-effort jobs low. Fined-grained resource allocation enables DARSS to schedule more tasks than a coarser-grained approach would. Furthermore, DARSS schedules production jobs as close to their deadlines as possible. This scheduling policy allows the system to maximize the number of low-priority tasks that can be scheduled opportunistically. DARSS is a scalable system that can be integrated with YARN.

DARSS is evaluated on a simulated cluster of 300 nodes against a workload derived from Google Borg's trace. DARSS is compared with Microsoft's Rayon and YARN's built-in scheduler. DARSS achieves better production job acceptance rate than both YARN and Rayon. The experiments show that all of the production jobs accepted by DARSS complete before their deadlines. Furthermore, DARSS has a higher number of best-effort jobs serviced than Rayon. And finally, DARSS has lower latency for best-effort jobs than Rayon.