CSR: Small: Collaborative Research: Enhancing Cloud Performance With On-Demand Isolation

University Of Pittsburgh, Pittsburgh PA

Investigators

Abstract

The modern trend in computing systems is towards system architectures containing a large numbers of heterogeneous computational and I/O resources. Unfortunately, while the increase in scale allows increased workload consolidation, wherein a single system runs multiple independent applications in parallel, it does come at the cost of introducing increased interference across the different application workloads. Workload interference is the result of the behavior of one application impacting the performance of another, even if both applications are running on different hardware resources. This can be due to contention on shared hardware resources (such as last level caches, memory controllers, or I/O devices) or even software resources managed by the operating system. Cross workload interference is especially problematic for large scale shared infrastructures such as cloud hosting services, which rely on co-hosting large numbers of widely disparate workloads inside a single datacenter environment. Preventing interference effects is critical for cloud computing to fully deliver on its promise as a universal computing substrate. This project addresses the problem of cross workload interference, by providing a holistic system that both detects the impact of interference on applications and mitigates its effects by providing dynamic isolation capabilities in the underlying system software. This approach relies on the ability to dynamically partition the underlying hardware resources such that isolation is achieved at the hardware layer, while also allowing the partitioning of system software to avoid contention on more abstract resources present in the system software itself. To achieve these goals this work implements a "Virtual Platform" abstraction representing an individual and isolatable system domain assigned to a particular task or workload and consisting of one or more virtual machine instances. The virtual platform itself is assigned an allocation of hardware resources consisting of independent "isolatable units." These units are created through the decomposition of local hardware resources into the finest grained subdivision of resources that can be both individually allocated and effectively isolated from the rest of the system. While providing partitioned hardware resources to a virtual platform provides hardware level isolation, it does not address interference generated by the system software. Avoiding software level interference is achieved by partitioning the system software itself through Multi-Stack Virtualization. Multi-stack virtualization allows multiple independent system software layers to co-exist on the same local system by restricting their managed resources to the set allocated to a virtual platform. Taken together this system provides full isolation capabilities at both the hardware and software layers.