QoS-Aware Admission Control in Heterogeneous Datacenters.
Christina Delimitrou, Nicholas Bambos, Christos Kozyrakis · 2013
Large-scale datacenters (DCs) host tens of thousands of diverse applications each day. The DC’s cluster manager must determine “where ” and “when ” an application should run, and “how many resources ” it should be allocated. This is challenging in cloud environments for several reasons, including workload interference, dynamic resource demands, and hardware platform heterogeneity. Recent work has addressed “where ” an application should be scheduled, namely, which of the tens of thousands of servers has the configuration and interference profile that preserves the workload’s QoS guarantees. However, significant challenges remain in terms of “when ” an application is scheduled. DC applications have diverse QoS constraints and resource requirements. Typically workloads in DCs are scheduled in admission order, provided all workloads have the same priority. This is suboptimal for two reasons. First, demanding applications, i.e., non interference-tolerant applications, can block non-demanding workloads that can tolerate resource contention. Second, when system resources are fully occupied, the applications are queued until the system exits the oversubscribed phase, and resume scheduling as the first servers become available, to minimize scheduling delays. This can induce significant performance degradations as these servers might not reflect the workload’s true resource needs. We present ARQ, a multi-class admission control protocol that leverages Paragon, a heterogeneity and interference-aware DC scheduler. ARQ divides applications in classes based on the quality of resources they need and queues them separately. This improves server utilization and system throughput, while maintaining perapplication QoS guarantees. To enforce timely scheduling, ARQ diverges workloads to a queue of lower resource quality, if no suitable server becomes available within the time window specified by the application’s QoS. In an oversubscribed scenario with 8,500 applications on 1,000 EC2 servers, ARQ bounds performance degradation to less than 10 % for 99 % of workloads, while significantly improving utilization. 1.