My current research focuses on developing stochastic analytic models for performance and availability prediction of cloud computing based services. 

Background:

Cloud computing is a model of Internet-based computing where users have on-demand network access to a shared pool of configurable computing resources (e.g., networks, servers, storage, applications, and services) that can be rapidly provisioned and released with minimal management effort or service provider interaction. Critical obstacles in using cloud based services are (un)availability, performance and cost unpredictability. Developing an analytic and modeling based tool will help cloud service providers to create a pricing policy for the cloud users, and to compute the cost of user Service Level Agreement (SLA) guarantees.

Key problems:

Cloud based systems are inherently large scale, highly distributed, almost always virtualized, and operate in automated shared environments. Performance and availability of such systems are affected by a large number of parameters including characteristics of the physical infrastructure, characteristics of the virtualization infrastructure, characteristics of automation tools used to manage the cloud system, and so on. Because of this, the internal systems can be operating at any one of a very large number of system states. Any naive modeling approach quickly runs into state explosion and/or intractable solution. Scale of the cloud makes the problems challenging in terms developing practical solutions.

Our approach:

We are developing a comprehensive and high fidelity modeling approach. Solution of the stated problems calls for the joint analysis of availability, performance, energy usage, cost and capacity requirements. Our approach of interacting stochastic models, with the initial models being a set of interacting Markov chains, holds the promise of being scalable, tractable and yet reflecting all the relevant parameters affecting Service Level Agreements.

Broad impact:

Outcome of the research will be beneficial to the design and analysis of cloud services/systems (such as IBM cloud) as well as to underlying virtualized data centers for the capacity planning, Service Level Agreement management, energy consumption handling, and overall optimization. Once developed and validated, a new management tool will be implemented using the developed models as building blocks. The tool will accomplish the following three tasks: (1) for a well-characterized cloud service, it will predict the long-term system capacity requirements and workload arrivals in an offline manner, (2) when limited information is available on system resources and workloads (such as traffic bursts), it will provide short term forecasts and online decisions to manage the dynamic behavior of the system, (3) the newly developed tool will be embedded into the cloud management software/platform to detect the bottlenecks of planning and provisioning strategies at runtime.