A Multithreaded Resolution of the Service Selection Problem Based on Domain Decomposition and Work Stealing

In this paper, we are interested in the service selection problem. In the variant that we consider, the problem's inputs consists of: (1) a services' composition described as an extended business process, (2) a set of concrete services that can be associated with the nodes of the services' composition. The objective is to decide the best mapping between concrete services and composition nodes such as to minimize a penalty function. We consider for its resolution a view that was poorly investigated until now. It consists of mapping the service selection onto the constraint satisfaction problem. Based on the know-how in the resolution of this latter problem, we propose a new parallel algorithm for the service selection problem in shared memory contexts. The parallelization of service selection was investigated in the past. However, we differ from existing works on three points. Firstly, we adopt a more classical representation of the services' composition and parallelize a distinct sequential algorithm. Secondly, our solution is more scalable because the creation of parallelism is not limited by the structure of the composition graph. Finally, we combine two well-investigated techniques for parallelization: the domain decomposition and the work stealing. The proposed parallel algorithm is based on a general framework for which multiple instantiations are possible. We also propose an evaluation of the different instantiations and analyze their speed-ups. The experimental results show that super-linear speed-ups can be reached with our parallel framework.