Abstract
Workload models are typically built based on user and application behavior in a system, limiting them to specific domains. Undoubtedly, such a practice creates a dilemma in a cloud computing (cloud) environment, where a wide range of heterogeneous applications are running and many users have access to these resources. The workload model in such an infrastructure must adapt to the evolution of the system configuration parameters, such as job load fluctuation. The aim of this work is to propose an approach that generates generic workload models (1) which are independent of user behavior and the applications running in the system, and can fit any workload domain and type, (2) model sharp workload variations that are most likely to appear in cloud environments, and (3) with high degree of fidelity with respect to observed data, within a short execution time. We propose two approaches for workload estimation, the first being a Hull-White and Genetic Algorithm (GA) combination, while the second is a Support Vector Regression (SVR) and Kalman-filter combination. Thorough experiments are conducted on real CPU and throughput datasets from virtualized IP Multimedia Subsystem (IMS), Web and cloud environments to study the efficiency of both propositions. The results show a higher accuracy for the Hull-White-GA approach with marginal overhead over the SVR-Kalman-Filter combination.
Highlights
W ITH the growing ubiquity of cloud computing technologies over the past decade, cloud providers and researchers have strived to design tools for evaluating and enhancing different Quality of Service (QoS) aspects of their systems, mainly performance, availability, reliability and power efficiency
The kernel function was set to Radial Basis Function (RBF) since it is more appropriate to use on nonlinear datasets, while C and Gamma values were fixed at 0.1
As for the Kalman filter, we considered the transition as an identity matrix, we assumed a vector of zero control input, and the noise measurement was of the state directly
Summary
W ITH the growing ubiquity of cloud computing technologies over the past decade, cloud providers and researchers have strived to design tools for evaluating and enhancing different Quality of Service (QoS) aspects of their systems, mainly performance, availability, reliability and power efficiency. The development of system management policies that support QoS is crucial The latter is quite challenging, as it must rely on evaluation tools which are capable of accurately representing the behavior of multiple attributes (e.g., CPU, RAM, throughput, network traffic) of cloud systems [1]. Workload models allow cloud providers to evaluate and simulate resource management policies aimed at enhancing their system QoS before they are deployed in full-scale production environments. The use of predictable workload data is preferred for building offline workload models where a user wishes to forecast future load, while non-predictable workload data is preferred for building online workload models in which there are variations such as sharp, critical spikes in the load, which might impact the system
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
