Model-Based System Management for Multi-Tiered Servers

om plem nt Figure 1.3: Structure of this thesis forcement or the delay for obtaining a system state snapshot, both of which may be large. This thesis shows that intuitively correct performance optimization approaches, such as dynamic programming based resource-packing, do not necessarily yield the best cost-optimizing behavior when accounting for resource allocation delays and the fact that resource schedulers at multiple tiers may make independent optimization choices. This thesis proposes slow calibration of resource allocations as an alternative resource allocation optimization approach with better performance characteristics. However, it also shows that the usefulness of resource allocation adaptation strongly depends on the delay of resource allocation enforcement and the delay of acquiring system snapshots. When delays are long and interarrival time distributions have large variances, online optimization generates no benefit over a static resource allocation. 1.4 Thesis Structure Chapter II discusses the basic assumptions of this thesis and the fundamental concepts of multi-tiered systems and analyzes the design of typical server applications that are used in multi-tiered scenarios. This design study provides the foundational model of multitiered computations, which is then used to build application-transparent workload-tracking and policing features in the following chapters. Chapter III presents Stateful Distributed Interposition (SDI). SDI is the general mech-