AWAIT: Efficient overload management for busy multi-tier web services under bursty workloads

Abstract

The problem of service differentiation and admission control in web services that utilize a multi-tier architecture is more challenging than in a single-tiered one, especially in the presence of bursty conditions, i.e., when arrivals of user web sessions to the system are characterized by temporal surges in their arrival intensities and demands. We demonstrate that classic techniques for a session based admission control that are triggered by threshold violations are ineffective under bursty workload conditions, as user-perceived performance metrics rapidly and dramatically deteriorate, inadvertently leading the system to reject requests from already accepted user sessions, resulting in business loss. Here, as a solution for service differentiation of accepted user sessions we promote a methodology that is based on blocking, i.e., when the system operates in overload, requests from accepted sessions are not rejected but are instead stored in a blocking queue that effectively acts as a waiting room. The requests in the blocking queue implicitly become of higher priority and are served immediately after load subsides. Residence in the blocking queue comes with a performance cost as blocking time adds to the perceived end-to-end user response time. We present a novel autonomic session based admission control policy, called AWAIT, that adaptively adjusts the capacity of the blocking queue as a function of workload burstiness in order to meet predefined user service level objectives while keeping the portion of aborted accepted sessions to a minimum. Detailed simulations illustrate the effectiveness of AWAIT under different workload burstiness profiles and therefore strongly argue for its effectiveness. © 2010 Springer-Verlag.