Home > Events > Oral Candidacy - Frederick Nwanganga

Oral Candidacy - Frederick Nwanganga

Start: 12/15/2017 at 2:30PM
End: 12/15/2017 at 5:00PM
Location: 100 Stinson Remick
Add to calendar:
iCal vCal

 Frederick C. Nwanganga

Oral Candidacy

December 15, 2017        2:30 pm        100 Stinson Remick

Advisor: Dr. Nitesh V. Chawla

Committee

Dr. Gregory Madey        Dr. Ronald Metoyer        Dr. Michael Chapple

Title:

Optimizing Workload Resource Allocation on Public Cloud Infrastructure: A Data-Driven Approach A Dissertation Proposal

Abstract:

As more organizations migrate their services from local data centers to the Cloud, the problem of allocating the appropriate amount of resources to meet workload demands while minimizing cost and maintaining Quality of Service (QoS) guarantees becomes critically important. With a pay-as-you-go cost model, the migration of technology services from on-premises data centers to the cloud often requires that organizations also make a corresponding shift in their infrastructure operations budget from a Capital Expenditure (CapEx) driven model to an Operational Expenditure (OpEx) driven model. In this new OpEx paradigm, the direct correlation between infrastructure costs and resource consumption becomes even more apparent, resulting in increased cost visibility and demand to maximize resource utilization while minimizing costs. Most cloud migration efforts employ a heuristic resource-to-workload allocation approach when determining how to allocate services to cloud compute instances. While this approach may be sufficient for the initial lift, it is a very inefficient strategy for ongoing data center operations in the cloud. It increases the risk of over-provisioning and under-provisioning which result in increased Total Cost of Ownership (TCO) and Service Level Agreement (SLA) violations, respectively. Rather than relying on heuristics for the allocation of workloads to cloud resources, this research proposes the adoption of an analytic strategy. The proposed approach requires an understanding of the characteristics and patterns of workloads within a cloud computing environment. It also requires the ability to learn from prior workload behavior and prescribe an optimal resource allocation scheme or optimization policy based on set objectives and constraints. Finally, the analytic approach requires the ability to develop simulation models which provide insight into how workload characteristics change as a result of variations in resource parameters. This is critical in order to understand the impact of proposed optimization policies so they can be validated before being implemented in real world environments. Workload modeling and resource optimization in cloud environments is challenging due to the abstraction and overhead presented by the virtualization layer, the lack of clarity with and absence of sufficient system-level tracelogs that correlate to performance, the complexity and variability of workloads and a lack of effective standards for workload classification and forecasting. The proposed research will develop a mechanism to capture and analyze cloud resource utilization tracelog data, model and classify workloads based on the performance data, define a distribution to represent the model (or class), set the parameters of the model using parameter estimation methods and Goodness of Fit (GoF) tests and finally develop a cost and resource utilization optimization framework formulated as an Integer Linear Programming (ILP) problem. The mechanism will receive as input, prior resource utilization data for workloads with dimensions describing CPU utilization, memory utilization, network IO, disk IO and disk performance. It will produce as output, a recommendation for the optimal allocation of each workload to a host in order to minimize cost and maximize resource utilization within defined constraints.