Automatic Scaling of Cloud Applications via Transparently Elasticizing Virtual Memory

This dissertation addresses the topic of how to achieve elasticity of an operating system so that networked resources in the form of remote memory and computation can be scaled up, down and out to meet the dynamic workloads of today’s cloud applications. This dissertation shows that it is feasible to modify the Linux operating system to achieve transparent elasticity by implementing four key primitives: stretching of a process’ virtual address space across the physical memory of networked nodes; fine-grained jumping of process execution across the set of networked nodes participating in the stretched address space; pushing of memory pages to different nodes to create islands of locality; pulling of remote memory pages to satisfy a local page fault. The dissertation further evaluates the overall ElasticOS prototype and shows the benefit of the system compared to networked swap, as well as performs an analysis of the performance of individual components of the system