System Architectures with Virtualized Resources in a Large-Scale Computing Infrastructure

While system virtualization provides a new vehicle of consolidating computer systems, simply virtualizing existing servers does not realize all of the benefits, resulting in inefficient resource usage and underperforming systems. In this thesis, we design key components for three different applications on large-scale virtualized infrastructures. Unlike physical servers, the amount of resources allocated to a virtual machine (VM) can be changed dynamically after its creation. We first design a web server cluster that uses a hash-based request distribution algorithm to improve scalability and locality. To handle uneven request distribution, resource allocations are adapted according to current demand. We also develop a scheme to improve fault-tolerance by retaining existing locality in surviving VMs as much as possible and transferring the workload on the failed node with the virtual resources that were allocated to the failed node. We then design a distributed proportional-share CPU scheduler for reducing the execution time of parallel computing programs by adjusting the CPU share allocation. The scheduler first infers the dependency between computing nodes by monitoring network packets and process state at each node. Excessive CPU shares are then identified and transferred in a distributed fashion. While some resources can be simply space-shared between VMs, contention for time-shared resources may result in additional delay that affects the dead reckoning accuracy in Internet game servers. Finally, we improve the virtual network interface design by allowing timestamps to be taken in the host OS and passed to a guest OS. Our evaluation result shows that reduced response time, better throughput, and more accurate timestamps can be achieved in a virtualized infrastructure that incorporates the design proposed in this thesis.