Peer-to-Peer (P2P) computing has attracted considerable attention m the last few years because of its ability to aggregate unused resources. One of the fundamental operations in P2P systems is resource discovery, which is responsible for finding the resources that the users request. Resource discovery is challenging because the large number of resources to be
discovered can be in the order of millions, and because resources can join and leave in an arbitrary way. However, current resource discovery mechanisms in unstructured P2P systems tend to be inefficient; they are mainly based on the idea of the flooding the network with queries that consume a lot of bandwidth.
We have developed an efficient resource discovery mechanism for P2P networks which we have called the Blackboard Resource Discovery Mechanism (BRDM). The BRDM algorithm is based on the blackboard architectural concept which was first coined by researchers in the field of Artificial Intelligence. In a blackboard system, a set of problem solving modules uses a central repository (called the blackboard) for all shared information. The contents of the blackboard are facts and hypotheses made by the system during the process of solving a specific problem. In P2P networks, we assumed that each node maintains a blackboard and the query messages exchanged between nodes result in information being stored about the resources distributed across the network in the nodes' blackboards.
We studied the efficiency of the BRDM algorithm in compaiison to other algorithms using simulated networks of up to 10000 peers and different distributions for que1ies and resources.
We find that our algorithm can outperform other existing P2P search algorithms in the
following ways: it finds more resources, it generates fewer query messages; it is more
adaptable in dynamic P2P networks; it protects the anonymity of resource locations; and it reduces bandwidth consumption.
We have developed a system called ParCop, which can take full advantage of the BRDM algorithm. ParCop is a decentralized P2P system that provides the tools which allow users to distribute scientific computations. The ParCop architecture is novel because it uses BRDM to efficiently utilize the computational resources of the peers distributed across the network.
ParCop follows the master/worker paradigm, in which a peer can become a master as well as a worker. Other major properties of ParCop are: efficient scheduling policies which minimize the processing time of applications in the system; fault tolerance; efficient caching of the computational results; and scalability.
We presented experiments that showed that a better speedup can be obtained by using
ParCop for applications that are characterized by a high computation to communication ratio.
Other experiments also showed that proposed scheduling policies enhance ParCop's
efficiency in utilizing computational resources and that ParCop has fault tolerant qualities.