Article Preview
Top1. Introduction
Cloud computing is new paradigm which aims to provide real-time scalable and configurable resources over network to end-users. The advantages of cloud computing like dynamic scalability, pay-as-you-go model, billed and measurable computational resources, negligible infrastructure set-up and maintenance cost makes service providers to pay attention towards customers demands. Recently cloud computing service providers observe tremendous variations in job arrival rate, request processing requirements, requested task executing period due to increasing popularity of services like e-commerce, online scientific applications, social networking, high performance workloads, analytic services and so on. It is of great significance to satisfy accepted quality of service (QoS) constraints in such scenarios to avoid risk of missing potential customers (Carlos Vazquez et. Al.). Uncertainty of customer needs and increasing resource and energy cost makes recent studies more focused on resource scheduling and management based on load balancing, dynamic resource selection, effective placement and energy conservation (Zhuo Tang et. Al)(Kumar, J., Saxena, D., Singh, A.K. et al).
Payment for computing resources by user is governed by customized contract between client and provider called service level agreement (SLA). Fulfilment of SLA is done by monitoring quality of service (QoS) performance indicators as response time, availability, throughput, reliability, task waiting probability and security. Out of these, Response time is significant indicator that has to be taken care of (Boyun Liu et. Al). When service response time exceeds threshold mentioned in SLA, provider dynamically adjust resources. This leads to increased maintenance cost, risk of resources under and over provisioning and hysteresis (Jun Guo et.al).
Predicting service component response time for next request may solve these issues. Knowledge of request response time in advance helps dynamic resource allocator to allocate resources at right time (Jun Guo et.al), (Ballani, Karagiannis, Rowstron, Costa & Jalaparti, 2019). Use of time series forecasting for user request response time prediction is effective tool, as prediction about system future state is computed from current and historical statistics of service component. The main objective of this research is to find suitable time series forecasting model which accurately predicts upcoming request response time, for each Data Center (DC) if selected for schedulingHsu, Matsuda & Matsuoka, 2018). Followed by selection of DC with minimum predicted response time to serve next request. The key contributions in this paper are:
- 1.
Presents system architectural framework and principles for prediction based DC selection in cloud computing environment.
- 2.
A novel mathematical modelling of system based on M/M/m queuing theory and investigates optimization function for DC selection to minimize overall system response time.
- 3.
Design and development of dynamic resource provisioning and allocation algorithm, which select DC for upcoming request based on result of optimization function.
- 4.
Propose novel Dynamic forecast selector forecasting model, which automatically select best forecasting model out of considered time series forecasting models over iterations results in minimum forecasting error.
- 5.
Through extensive experiments find out suitable time series forecasting model for proposed architecture and optimize number of factors that affect forecasting accuracy such as model training set length, request arrival rate, type of workload and effective accuracy calculation method.
The rest of paper is as follows. Next section gives brief explanation of literature survey. Section three explains proposed system architecture and its components. Section four give details about forecasting models considered in this work, accuracy calculation methods and model training-testing dataset. Section five describes system queuing model, formulation of optimization function and proposed DC selection algorithm. Section 6 & 7 clarifies evaluation techniques and results obtained; finally section eight concludes the work.