Workload Modeling for Forecasting and Scalable Resource Provisioning in Cloud Networks
Abstract
Effective provisioning in the cloud demands an understanding of application workload dynamics. Current methods are extremely challenged given the mix and diversity of workloads and their on-demand deployment. This dissertation examines cloud workloads through statistical analysis to determine their salient features. These are employed to address three important issues in cloud environments: 1) Isolation of statistical models that accurately capture workload dynamics 2) Forecasting such workloads over short and long timescales, and 3) Resource provisioning to meet workload Quality-of-Service (QoS) requirements.
A diverse set of traces from production cloud environments was analyzed. Henceforth a methodology was developed, one that classifies cloud workloads according to their distinguishing statistical features. This methodology has enabled the isolation of a novel statistical model that describes salient features specific to cloud storage workloads. The model employs current methods from econometrics described as Autoregressive Conditional Score (ACS) in this realization. The forecasting algorithm realized from this model improves accuracy over existing methods, reducing forecasting errors by 10% up to 25%. The algorithm has been realized in MATLAB and integrated into the R statistical computing package. Furthermore, this model has inspired the development of a novel measure of workload burstiness, its variance, one that extends current measures by employing statistical features specific to each workload. That is, if a workload demonstrates standard or nonstandard variance properties, the burstiness is calculated accordingly. It is described as the Normalized Score Index (NSI). It also provides a means for burstiness comparison which facilitates resource provisioning given its [0,1] range expressible as a percentage. Thus workloads with scores of 0.2 and 0.6 are assigned resources by their scores.
Regarding resource provisioning, two novel algorithms have been realized each integrated as rate-based solutions applicable at integral points in cloud networks. The first targets the cloud network edge using the NSI in bandwidth provisioning. This was realized as a rate-adaptive framework applicable in virtualized cloud environments. The second algorithm is developed specifically as a scalable solution given enterprise cloud infrastructure. The methods have been tested under diverse conditions and found to provide tractable bandwidth assurance and foster QoS provisioning in dynamic cloud environments.
Description
Keywords
Physics--Theory, Statistics, Engineering--Electronics and Electrical
Citation
Adegboyega, A. (2017). Workload Modeling for Forecasting and Scalable Resource Provisioning in Cloud Networks (Doctoral thesis, University of Calgary, Calgary, Canada). Retrieved from https://prism.ucalgary.ca. doi:10.11575/PRISM/28723