Multi - tier Internet Service Management : Statistical Learning Approaches

Modern Internet services employ a multi-tier architecture. While facilitating flexible service deployment, the tiered architectures introduce several significant challenges. Complex inter-tier dependencies and dynamic bottleneck tier shift are challenges inherent to multi-tier architectures. Highly dynamics of session-oriented Internet workloads further magnifies the complexity of multi-tier Internet services. Managing these complex services for session-oriented performance improvement and quality of service provisioning is an important but challenging undertaking. Three critical Internet service management mechanisms are: admission control, dynamic resource provisioning and service differentiation. Significant research literature studies the three mechanisms using a variety of techniques. Queuing based analytical models of the multi-tier systems have been proposed to achieve request-oriented performance guarantees. However, queuing models often lack the ability to capture session workload dynamics in multi-tier systems and queueing-theoretic approaches are often not effective in resource management for quality-of-service provisioning in highly complex multi-tier systems with highly complex session-based workloads. This thesis concentrates on applying statistical learning based approaches for multi-tier Internet service management to achieve efficient, balanced and scalable services. The main advantage of statistical learning approaches is that they solve complex dynamic problems through learning and adaptation, and require no priori domain-specific knowledge. First, we will explore a session based admission control strategy to improve session throughput in multi-tier Internet services. Using a bayesian network model, the approach will achieve coordination among multiple tiers of the service for effective session throughput improvement. Second, we will promote the use of a novel relative session-oriented performance metric, session slowdown. Session slowdown represents the user-perceived performance of a session. We will further propose a statistical regression based dynamic resource provisioning strategy for session slowdown guarantee in multi-tier systems. While offline training captures the dynamic behavior of the service as statistical regression models, online monitoring utilizes the learned models to predict the multi-tier Internet service’s resource requirements. Third, we want to design a service differentiation approach that employs a combination of reinforcement learning and neural networks to achieve session slowdown based differentiation in multi-tier Internet services. The approach aims to adaptively achieve both absolute and relative differentiation among multiple customer classes. Finally, we will develop a user interface based Multi-tier Internet Service Management Console, which can be used by a system administrator to monitor and finely tune the service performance. Successful completion of this research will guide using statistical learning approaches for quality-of-service provisioning in modern multi-tier Internet services.