CSR:Medium:Collaborative Research: An Analytical Approach to Quantifying Availability (AQUA) for Cloud Resource Provisioning and Allocation

Suny At Buffalo, Amherst NY

Investigators

Abstract

Cloud computing will significantly transform the landscape of the IT industry and also impact the economy and society in many ways. The reliability and availability of cloud services, affected by various hardware and software component failures, becomes increasingly more critical, as government agencies, business and people are expected to rely more and more on these services. Lack of a guaranteed high availability of cloud services and applications is considered by many IT professionals as the top concern for preventing a successful implementation of cloud services, followed by device based security and cloud application performance. This project aims to predict the service availability for a given setting, and design effective resource provisioning and allocation algorithms to guarantee a high availability level required by cloud services. The project is expected to significantly advance the state-of-the-art by offering deep insights into the knowledge about accurate prediction and cost-effective guarantee of availability/reliability of cloud services. The outputs from this project can be used to 1) improve service availability, performance and resource utilization while minimizing the cost of overprovisioning, 2) reduce huge losses in revenue and productivity due to service outages while enabling new (mission-critical) applications and services. The existing approaches to ensuring availability are qualitative in that they use heuristics to duplicate data or restrict the number of virtual machines (VMs) that should be placed in the same rack/server to improve reliability/availability of cloud services. However, it is essential to be able to quantify availability for a given setting. Quantifying availability for an often finite service duration via analysis (as opposed to measurement or qualitative evaluation) requires transient, instead of steady state probability analysis based on a wide range of failure and repair/backup models. This project takes a holistic approach to addressing the open challenges via both rigorous analysis and extensive experiments. More specifically, the project leverages two large-scale HPC/Cloud production systems at PI?s institutions to generate a rich set of fine-grained data about physical component failures (which is not available in the public domain). The data is then analyzed to build and verify/validate failure models. Based on the failure models and for a given Infrastructure-as-a-Service (IaaS) request for n virtual machines (VMs), a service duration of t time units and a desired availability level a < 1, the project develops an analytical model to predict the availability that can be achieved for the service duration (t), if an additional k backup VMs are allocated. The project also develops cost-effective, multi-objective optimization based cloud resource provisioning and allocation algorithms that determine the appropriate value for k (and the placement of these n+k VMs) in order to achieve the required availability level a.