Research On High Performance Multi-networking Communication In Parallel System

Since 1990s, parallel computing has developed unprecedented rapidly, structure of parallel computer system has matured, data transmission network has become standardization and transmission rate has lifted , which all have created advantaged conditions for the development of parallel system. Cluster systems have become the main achievements of parallel computing. As the performance of node's calculation improving continuously, communication system in cluser has become an important factor for parallel computing performance. To satisfy cluster's demand for performance, reliability, maintainability and other requirements, it is necessary to improve the communication system's performance. The new high-speed networks are usually expensive, while the traditional low-speed networks are not fully utilized because they are not meeting the needs of high-performance communication. It is necessary to take the idea of parallel processing to make the best use of available communication resources. Multi-networking communication(MNC) technology focuses on exploiting multiple networks to improve communication performance and reliability.This paper is one important part of Liaoning Science Technology Fund Doctor Startup Project: Research on High Performance Multi-networking Communication in Parallel System(No. 20051058). This mainly studies communication system of parallel system, the main content includes:1. It analyzes the major components of the communication system, the basic theory of MPI, which is the most important parallel computing programming tool, and the importance of that the communication system to the whole cluster performance.2. Basing on the codes of MPICH-1.2.5 and Linux-2.4.32, it specific analyses the MPICH's communication process, which including three layers: application layer, kernel layer and NIC layer. From the send and receive functions of MPICH layer to the process of socket and TCP protocol in kernel layer, and finally the NIC transmits data between the nodes and network. And it expatiates importance of communication system on parallel computing performance from three sides, gives the approaches how to improve the communication performance.3.On the existing experimental environment, this paper proposes the optimization lightest load selection arithmetic(OLLS), which is basic on the bonding technology. This arithmetic generates an index according to the NIC's current load status, which shorts the time for finding the lightest load NIC, realizes the dynamic choice of networking communication, optimizes the performance of the entire system. Finally uses the MPBench testing the broadcast communication, bandwidth and roundtrip, the results show that the optimization lightest load selection arithmetic is better than bonding technology and one-networking communication.