S.V. Kovalchuk - Ph.D. (Eng.), Senior Research Scientist, National Research University of Information Technologies, Mechanics and Optics
A.M. Chirkin - Student, National Research University of Information Technologies, Mechanics and Optics
K.V. Knyazkov - Ph.D. (Eng.), Senior Research Scientist, National Research University of Information Technologies, Mechanics and Optics

The development of the second-generation cloud computing technologies based on the AaaS (Application-as-a-Service) model discover new requirements to the methods of resource characteristics estimation, which are needed to obtain the results with the required quality within the defined time period. These requirements are defined by the fact, that within the AaaS model the user can operate with composite application containing blocks which significantly differ by the execution time depending on the resources being used. This requirement distinguish the AaaS model both form IaaS or PaaS models (where resource requirements are predefined by the user) and from SaaS model (where resources are couples with the particular software offered within the cloud computing environment). The key metric for resource assessment here is the execution time for different resources. This work describes the approach, which allows: a) performance modelling and execution time prediction for composite applications; b) models adoption according to the dynamically changing characteristics of the infrastructure; c) uncertainty estimation for the predicted execution time of composite application. The work define a set of factors that have an influence on the execution time of services and composite application within the cloud computing environment. Each of these factors is characterized by the individual specificity of variation. Moreover they are distinguished by the impact into the total application execution time. Nevertheless, each of them can be generally considered as random variables, optionally depended on the parameters of the following groups: algorithm characteristics for the software (or set of software in case of composite application), resource characteristics, input data characteristics and platform characteristics. The sum of these random variables can be considered as an estimation of the execution time. Two main classes of cloud computing performance models are considered: parametric models, which formally depend on the input data parameters, and stochastic models, which are characterized by the distribution invariant for particular service (the distribution could changes in case of the service state is changing, but still is independent to the input data of particular task). As alternative solutions the following models are used: a) models built automatically using machine learning algorithms applied to the execution history; b) explicitly formalized models which describe the software available within the cloud computing environment. The work presents the unified approach which allows to estimate the execution time of cloud computing services and composite applications and support the cloud computing environment management for the services planning and resource cost estimation. The solution developed based on iPSE concept and CLAVIRE platform enables automation of the performance models identification, validation, correction and usage processes. One of the features of the presented solution is ability of estimation and re-estimation (rectification) of the execution time dynamically as the execution statistics is collected during the software usage.