并行系统高性能通信软件若干关键技术的研究

消息传递层的有效实现、层间通信性能损失问题的解决以及ＭＰＩ集群通信的高效实现对提高用户应用层网络性能具有重要意义。本项目拟研究新的消息传递方法和机制，建造低通信延迟和高有效带宽的底层库通信软件，建立新的ＮＦＭ与ＭＰＩ层的高效接口，研究一种新的ＭＰＩ集群通信技术，以减少ＮＦＭ与ＭＰＩ接口间性能损失，从而使网络通信性能明显地得到提高。

In order to reduce the gap between Cluster and MPP, we studied the methods and mechanisms for effective message passing, and implemented the Fast Message layer (FM), which provided services for application layer. Usually, performance of the network in cluster is greatly reduced by the ineffective interface between message passing layer and MPI higher library layer, therefore, only when the higher MPI parallel communication mechanisms are improved, can those higher applications benefit from the fast network hardware and effective communication software. At present, the cluster communication of MPI is based on point-to-point mechanism, so the communication performance is very low. This drawback not only reduces the scalability of algorithms and whole performance of applications, but also limits the benefit of high-performance of the lower communication system. So the effective implementation of message passing layer, the solution of communication performance losing problem between layers and the resultful improvement of MPI cluster communication mechanisms are all important for improving the network performance of application layer..The main purpose of this project is to study new method and mechanism for message passing, build low-latency and high-bandwidth communication software in low layer library, implement high-effective interface between new fast message passing layer (NFM) and MPI layer, design a new MPI cluster communication mechanism in order to reduce the performance losing between NFM and MPI interface and thereby greatly improve the network communication performance..Several important results are achieved through detailed study:.1).An effective NFM layer is designed based on SMP-cluster connected with Myrinet, which greatly reduces the cost of memory copy, simplifies the buffer management, and implements reliable, orderly communication and flow control in low cost..2).The interface between NFM layer and ADI is optimized, which greatly enhances the whole performance of MPI-NFM, because the communication layer need not wait for the whole message before transmitting it any longer. .3).New techniques of MPI cluster communication are designed and a series of global communication algorithms are brought forward, which improve the COLLECTIVE communication performance in MPI and achieve good results in both experiments and applications..4).The influence of multi-thread reusable network to communication performance on Myrinet/GM network platform is studied. A dynamic strategy based on network interface layer for the selection and distribution of communication link is raised. And a new protocol layer MNC is designed and implemented, which supports the parallel communication of multiple Myrinet networks, maintains the equality of communication processes, makes most use of the multiple Myrinet network resources. So both the effective bandwidth of a single link channel and the communication performance of application layer are enhanced. All of the improvements are effective in practical applications. .