The object sharing space is often used as a technology to realize distributed parallel processing that can compute computationally intensive and time-consuming problems in a short time. However, since the model is a server-client model, accesses are concentrated on the server that provides the object sharing space. Therefore, the more clients there are, the more bottlenecks occur in network communication, which hinders the performance improvement of distributed parallel processing. Therefore, the number effect, which is a performance indicator in distributed parallel processing, decreases as the number of clients increases. In this study, we developed a distributed parallel processing system using P2P, which allows clients to communicate with each other. By implementing unicast communication between clients for sending and receiving information, we have achieved distributed parallel processing using P2P. In addition, the main processing of the system, such as P2P communication, is implemented in a software structure that is independent of the specific computational processing. As a result of the evaluation, we confirmed that the performance of the distributed parallel processing was improved compared to the existing method using the object sharing space. Furthermore, we confirmed that the number effect does not decrease even when the number of clients increases. In addition, it is possible to implement distributed parallel processing with fewer files and lines than before.
View full abstract