"Must-Work": A Scalable Model for Parallel Recursive Problems on P2P Networks

Most of real world problems are cpu-time intensive; their solutions take too much time to compute using a single machine. The grid and cloud computing offer a potential solution to this problem. However, such solutions are in general expensive. An alternative solution uses P2P networks, a set of machines in the Internet which collaborate to perform the same task. Branch-and-Bound is a model of such solution, but most of parallel applications developed on P2P networks are based on the Master Worker paradigm, particularly for divide and conquer problems (D&C), where the Master divides the tasks and sends them to Workers, while the Workers execute received tasks as in the client server model. This solution suffers from the scalability problem, and, as a consequence, hierarchical Master Worker model was introduced. Scalability is improved, but it still remains a critical issue. In this paper, we propose a new generic and scalable model for parallelizing the resolution of the recursive problem (a type of divide and conquer problems) on an existing P2P architecture. As opposed to the existing hierarchical Master-Worker models, in our solution, each network node is both a Master and Worker called "Must-Work". The proposed solution uses a dynamic tree for tasks distribution; it is constructed according to a node requestor. We have evaluated our solution using The Quicksort method under the MPI platform. The results are globally satisfactory in term of time execution compared to the sequential solution.