Today, we are seeing the beginnings of a new way of developing and delivering IT services. On the one hand we have the emergence of grid and cloud computing. On the other, we are witnessing the first attempts to replace the forty year old practice of buying monolithic software solutions with new strategies in which solutions are exactly dimensioned to meet specific needs and provide only the functionality that is actually required, and in which resources are purchased only as needed. At least within companies, Service Oriented Architectures make it possible to compose software flexibly. The initial attempts point to a new way of delivering IT-services. This is the theme of the Collaborative Research Centre 901 “On-The-Fly Computing“ (CRC 901).
In the vision pursued by the CRC, On-The-Fly (OTF) Computing consists of the flexible composition and delivery of IT-services, that are traded freely on the market. In the longer term, the goal is to support transparent interactions between hard and software from different suppliers with a minimum of human intervention. In parallel with these developments, it will be necessary to support the development of a vibrant market for OTF services, maintained through transactions between market players. The CRC 901 is currently investigating the feasibility of such a vision, developing methods and techniques that provide broad-ranging capabilities to automatically configure, execute and adapt IT-services available on global markets.
Given the normal level of granularity offered by OTF markets, it is often possible to respond to a set of user requirements with several functionally equivalent or very similar services (see sketch in figure). Usually, such services are differentiated by non-functional properties such as response times, cost and memory requirements. Once these non-functional properties and the user’s requirements have been well defined, a quality-based evaluation can identify the best possible combination. C-LAB is currently working on procedures to automate the evaluation and selection process, both in static and in highly dynamic environments. This work is part of its contribution to CRC 901 sub-project B2 “Configuration and Rating”.
The first step is to analyze and systematically identify the components of the system, including functions to aggregate non-functional properties for a composite IT-Service, optimization procedures to handle restrictive user constraints, and utility functions to reduce different types of non-functional properties to a small set of numerical values. Building on this analysis, the study is investigating the feasibility of Reinforcement Learning (RL) as a solution for OTF computing. Standard RL techniques such as Q-Learning or SARSA allow agents to adapt to a dynamic environment and hold much promise as a way of handling the continuously fluctuating supply of services on OFT markets. But before they can be used it will first be necessary to transfer the associated decision making problems from the OTF domain to the RL domain, using an appropriate model. C-LAB’s current research is addressing these challenges.