Model-based system management for multi-tiered servers.

Internet services are increasingly designed as multi-tiered, i.e., composite services linked by a fast and reliable communication network infrastructure. Even though multi-tiered systems are common, their problems are only insufficiently addressed by current operating systems (OS) designs and existing literature. This thesis specifically addresses performance management problems in multi-tiered server deployments and proposes a generalized framework which facilitates system-level management of multi-tiered activities as an add-on for standard OSs. The problem of interference between co-located multi-tiered services is addressed by proposing Virtual Services (VSs) to control the performance of shared back-end services. Since the configuration of VSs requires at least a model-based understanding of the multi-tiered system, Performance Maps (PMaps) is introduced. PMaps infer the dependencies between interacting services in a multi-tiered system through observation and online service model understanding. This thesis demonstrates that metrics based on PMaps can be used to identify several performance problems that occur in multi-tiered setups. This thesis also addresses the question of whether it would be beneficial to integrate VSs and PMaps into an online resource allocation adaptation approach that attempts to optimize an external cost or utility function by changing resource allocations. Two calibration-based approaches are proposed and shown to perform nearly as well as, or better than, aggressive rescheduling of system resources. However, it is also shown that resource allocation enforcement delays negatively affect the performance of online resource allocation adaptation, thus limiting its usefulness. This commonalities between VSs and PMaps are captured by the proposed system support layer called Stateful Distributed Interposition (SDI). It is designed to simplify the adaptation of single-host systems for their use in multi-tiered setups. SDI supports the addition of arbitrary state to OS entities without requiring any kernel recompilation. SDI automatically propagates this state according to system-specific propagation rules (alongside multi-tiered activities) from one tier to another. This attached state can be used to trigger and control OS plugins, such as PMaps and VSs. A prototype of SDI is implemented and shown to add only slight (