Parallel processing over a peer-to-peer network : constructing the poor man’s supercomputer

Abstract

The aggregation of typical home computers through a peer-to-peer (P2P) framework over the Internet would yield a virtual supercomputer of unmatched processing power, 95% of which is presently being left unutilized. However, the global community appears to be still hesitant at tapping into the well of unharnessed potential offered by exploiting distributed computing. Reasons include the lack of personal incentive for participants, and the high degree of expertise required from application developers. Our vision is to tackle the aforementioned obstacles by building a P2P system capable of deploying user-defined tasks onto the network for distributed execution. Users would only be expected to write standard concurrent code accessing our application programming interface, and may rely on the system to transparently provide for optimal task distribution, process migration, message delivery, global state, fault tolerance, and recovery. Strong mobility during process migration is achieved by pre-processing the source code. Our results indicate that near-linear efficiencies – approximately 94% ± 2% of the optimal – may be obtained for adequately coarse-grained applications, even when deployed on a heterogeneous net- work.