Analysis of control-theoretic predictive strategies for the adaptation of distributed parallel computations.

ABSTRACTIn adaptive distributed parallel applications the adaptation process is based on the ability to change some characteristics of parallel components, such as the parallelism form and the parallelism degree, in response to unexpected execution conditions. Although existing research work has studied this problem, it is of increasing importance to investigate adaptation strategies able to reach important properties like the stability of control decisions, i.e. to guarantee that reconfigurations are effective and durable, and control optimality, expressed by means of cooperative and non-cooperative agreements between decisions of different controllers. These properties are crucial in distributed environments like Grids and Clouds, where reconfigurations imply a cost both in terms of a performance degradation as well as a monetary charge. In this paper we briefly introduce the basic ideas of our methodology and we introduce different adaptation strategies based on alternative formulations of the Model-based Predictive Control technique. First hints about the effectiveness of our approach are discussed through experiments developed in a simulation environment.