AUS1

Dynamic adaptation and self management of compositional systems

We assert that autonomicity will be a key component of the software architecture and realization of these complex systems. Primarily because there is unlikely to be centrally mandated goals or levels of service, any universally-agreed protocols or other technical standards, and there will be dramatically reduced a priori knowledge of the topology or component population in any given system. Specifically, we believe that next-generation systems must be designed to be radically distributed and decentralised, self-describing, self-organising, self-managing, self-configuring and self-optimising. And that these systems must provide a seamless communications infrastructure composed of multiple technologies and be able to leverage local information and decisions without sacrificing global performance, robustness and trustworthiness. Our challenge is to understand the design and operational trade-offs required to ensure systems can operate correctly within their specified parameters; and also be able to adapt to a changing environment while maintaining this consistent behaviour. In AUS we are specifically focusing on the dynamics of such self managing systems. The key project goal for the (single) project in AUS is to advance the state of the art in Dynamic adaptation and self-management of compositional systems. The focus is to be on Tools and techniques to ensure that self-managing software behaves predictably at runtime. The key research questions are:

How do we describe the envelope of acceptable behaviour?
Can we use visualisation as a tool to facilitate this description?
How can we optimise runtime behaviour to ensure conformance to this operational-envelope?

The key results anticipated by the end of the project are:

Review of autonomic and self-organisation techniques (complete)
Positioning autonomic computing for automotive industry (complete)
Development of experimental software base (complete)
Model-based approaches to describing adaptive behaviours (current)
Visual analytic metrics to describe software and sensor data navigation, aggregation and exploration (current)
Novel decentralised collaborative optimisation techniques for autonomic components (current)

Overview of achievements in the reporting period

We have sucessfully demonstrated the application of dynamical systems techniques to (simple) network systems, and have begun applying the same approaches to sensor networks in conjunction with the Clarity CSET (Dr Dobson).

We have applied a slightly different suite of techniques deriving from Bayesian network analysis to fault detection and diagnosis (Dr Gaudin). This research yielded a significant speed-up in the ability of a client to develop fault models that can then be used to aid diagnosis in a data centre environment.

We have deepened our involvement in autonomic networks by investigating the effects of topology on key properties such as information propagation (Dr Cellai). This will inform future work on adaptive networks in which the adaptations are guaranteed to preserve critical properties.

Plans

Initial experiments have proved promising, but require extension in three key respects.

Firstly, we need to deepen and broaden the case studies attempted. We have begun this process by exploring the application of dynamical systems analysis to re-enforcement learning (Lero@UCD and Lero@TCD). Applications in sensor networks are also being pursued, both in collaboration with Clarity (in environmental wireless sensor networks) and in collaboration with Rutgers University in New Jersey that will provide a platform for further research.

Secondly, the mathematical techniques being deployed need to be expanded. Specifically we are interested in linking the specification of autonomic control models with environmental models (such as the network load, isochrony requirements etc). Such linkages are essential in developing well-founded adaptive control, since they give confidence that the systems being developed take account, at a fundamental level, of the environments into which they are deployed. We are also introducing elements of more classical specification and analysis and elements of network science. At a practical level we are seeking to engage with other industrial partners in the area of fault diagnosis and enterprise systems management (led by Dr. Gaudin).

Thirdly, we believe that it is essential to look more fully at programmatic aspects, in terms of how correct adaptive systems can be developed and deployed. We have begun the process of starting another AUS project in this area managed by Dr Butterfield and to be led by a new post-doctoral researcher. Additionally, the Automonic System Specification Language (ASSL) is being extended, supported by tools, and applied in various areas as a means of providing formal development support for autonomic systems (led by Dr. Vassev)