[Back]

P. Patonai:
"Application usage Profiling and Forecasting in Shared Cloud Systems";
Supervisor: I. Brandic; Institut für Informationssysteme, Distributed Systems Group, 2014; final examination: 2014-10-07.

This thesis addresses the problem of elastic cloud resource scaling. Cloud providers have to offer on-demand resource provision for time varying workloads. This often leads to SLA violations on application level, or energy waste due to providing more physical resources for the hosted applications than needed. Providing always as much physical resources as needed, results in QoS stability for the hosted applications. Also, significant reduction of electricity consumption
can be achieved by dynamically turning down unused resources. This makes the cloud providers more competitive.
Thereby we describe an adaptive and scalable scheduling infrastructure. This monitors the cloud applications, selects automatically an appropriate statistical forecasting method. Then,
based on the forecast values, optimizes the resource allocation in the cloud. The contribution of this thesis is three-fold: (1) It presents a detailed survey about the available forecasting and time
series analysis tools. (2) It provides a flexible and automatic classification framework to choose the best fitting forecasting method, based on the features of the monitored traces, such as trend,
seasonality, and variance. (3) In the course of the work, a cloud simulator is developed in order to compare different forecasting methods, how accurately they can predict SLA violations.
During the work it was observed that a high proportion of cloud applications are exhibiting daily periodical fluctuations in terms of resource needs. Thus, we are focusing on seasonal forecasting,by applying the Fourier transformation based forecasting, the Holt-Winters exponential smoothing, the neural network autoregression, and the STL decomposition based forecasting.
These seasonal forecasting tools are evaluated by using the implemented cloud simulator. For the test we use synthetically generated traces of applications with periodically fluctuating cpu,
memory, network, and disk I/O usage. We compare the used forecasting methods by the number of correct, incorrect and false positive predictions of SLA violations. In relation to the simulation
we discuss additional aspects, such as heuristic based target machine selection and migration, the usage of forecast values to create scheduling rules, and the adaptive error corrections. The ultimate aim of the simulator is to test the interaction of various approaches and the effect of those approaches on cloud utilization and on violation detection.