|Thesis abstract: |
Datacenter costs and power consumption are rapidly growing. Overprovisioning resources and running servers at low utilization is no more affordable. In recent years, the availability of virtualization technologies on commodity server hardware provided the opportunity to consolidate workloads on a smaller number of servers without introducing compatibility and security issues. Server consolidation reduces the power consumption and the operating costs of the datacenter, but introduces a number of challenges. In fact, as server utilization increases, building and parameterizing accurate models of workloads is more important than ever to predict the response times of the systems. At the same time, the opportunity to relocate workloads leads to interesting research questions regarding the optimal mapping of workloads to servers. While server consolidation is a static workload placement problem, dynamic workload placement problems occur in Infrastructure-as-a-Service (IaaS) clouds, where a scheduler needs to decide the optimal allocation of a newly provisioned virtual machine. In this dissertation we present the following contributions, related to both optimal workload placement and performance model parameter estimation: - A cooling-aware static workload placement problem in which the total datacenter power, including server and cooling power, is minimized, while satisfying constraints on the response times. A mathematical programming formulation of the problem as well as efficient heuristics are proposed and validated. - A dynamic workload placement problem for IaaS clouds, in which the profit from running virtual machines on a set of heterogenous server is maximized. An optimized tool to compute the optimal scheduling policy was developed. - An algorithm to estimate the service demands, a fundamental parameter of queueing network models, when the system is subject to changes (e.g., the modification of virtual machine monitor parameters). The algorithm successfully estimates the parameters under the different system configurations. - Effective strategies to estimate the parameters of a class of stochastic processes, Marked Markovian Arrival Processes (MMAPs), which allow to study the performance of correlated arrivals at queueing systems. We propose the first techniques to estimate the parameters of MMAPs based on closed analytical expressions.