Performance Centric Architecture & Prototype of Portable File System for Virtual Environment

Abstract

In this era of cloud computing, the focus is shifting on providing resources from hardware infrastructure to software applications on demand. This environment has two major advantages; relief from con guring or setting up the environment and other being resource scalability, i.e. utilizing required quantum of resource. Popularity of these features has motivated developers to convert di erent components into on-demand paradigms. File systems are one such components which have not been widely transformed into on-demand models. Some service providers have come up with their own le systems dedicated for cloud based storages (Google File System, AWS S3, etc.), but there is a long list of kernel le systems and native NFS/DFS clients which are not adapted to this environment. These systems have to be explicitly con gured by the user in their virtual machines. Con guring them is not a trivial task and requires tons of dependencies to be resolved leading to a hectic schedule. In cloud environment, this task becomes more complicated when the guest OS is either a customized image or has restrictions imposed by the service provider. Even if these shortcomings are addressed in cloud, the user still has to follow the tedious task of con guration which clearly indicates that le systems are not utilizing on demand paradigms of cloud computing. This has resulted into user settling to default le systems already available within the operating system immaterial of whether it would be suitable for their speci c workload scenarios. Optimized storage & memory do not su ce optimal results if the le systems with required parameters are not chosen correctly. Kernel le systems and native NFS/DFS clients are dependent on underlying OS architecture and rigid towards kernel versioning. In order to make their adaptability as on-demand models, i.e. providing le systems in form of service, this dependency needs to be addressed. This research work describes Portable File System Framework which addresses the issue by making them portable. This framework runs le system outside user's environment in a separate light weight virtual machine termed as "Engine", and ports it to user's OS via abstraction layer. This layer maintains le handler semantics giving feel as if the le system is running in user's environment. The portable framework has been developed on the doctrines of User Space File Systems, which have gained huge popularity over recent years. This work creates an ecosystem of providing an independent identity to le systems giving the scope of attaching/detaching them on the y. The framework has been designed to be exible enough to get integrated with OpenStack cloud manager (or any other manager) to provide le systems on demand in form of service. The framework has been both micro and macro benchmarked along with trace analysis to examine overheads. A comparative of the kernel version with the portable counterpart is drawn as an inference for each analysis. To accentuate the potential this ecosystem will provide, 5 distinct case studies have been presented which has helped in solving long pending issues related to kernel le systems & clients.