Coordinated session-based admission control with statistical learning for multi-tier internet applications

Popular Internet applications deploy a multi-tier architecture, with each tier provisioning a certain functionality to its preceding tier. In this paper, we address the challenging session-based admission control issue for multi-tier Internet applications. The session-based admission control approach (SBAC) designed for a single Web server is not effective for a multi-tier architecture. This is due to the fact that the bottleneck in a multi-tier website dynamically shifts among tiers as client access patterns change. Admission control based on only the bottleneck tier is not efficient as different sessions impose different resource consumptions at the different tiers. First, we propose a multi-tier measurement based admission control (MBAC), which pro-actively accepts different session mixes based on the utilization state of all tiers. More importantly, we design a coordinated session-based admission control approach (CoSAC) based on a machine learning technique. It uses a Bayesian network to correlate the states of all tiers. The probability with which a session is admitted is determined by the probabilistic inference of the network after applying the evidence in terms of utilization and processing time at each tier to the network. We compare CoSAC with MBAC and a black-box approach tailored from SBAC, using the industry standard TPCW benchmark in a typical three-tier e-commerce website. Experimental results demonstrate the superior performance of CoSAC with respect to the effective session throughput.