Networked parallel computing system NPCS (sometimes termed as cluster computing system) has become increasingly important in high performance computing. This paper describes a networked parallel supercomputer THNPSC-1 built at Tsinghua University. It consists of 16 dual pentium III SMP nodes. It employs two kinds of high speed interconnection networks. One is custom designed crossbar switch network THNet with dynamic arbitration and routing capability, built by Tsinghua University. The other is a 100Mbps Ethernet. The crossbar switch THSwitch of the THNet is built on the basis of Altera FPGA chips with 150,000 gates, and THNet also contains the network adapters THNIA. THNet is able to provide 1.056 Gbps data transmission rate for each port, and the aggregate bandwidth can reach 8.4 Gbps. By using the technique of fixed user message buffer management and extend active message［9］, THNet carried out the user level message passing, bypassing the operation system with zero data copy. The measurement indicates that communication latency of message passing can be reduced to 8 Microsecond. The software of THNet includes the driver of THNIA, and the library for supporting the user-level communication. This paper also gives a comparison with the related work, and depicts the applications of THNPSC-1.%网络并行计算（也称集群式计算）是实现高性能计算的重要方式，该文介绍了一个清华大学研制的网络并行超级计算系统THNPSC-1.它是由Pentium III SMP计算结点组成；网络互联采用两种高速网：一种是自制的具有动态仲裁与路由寻经的交叉开关网络THNet,另一种是100Mpbs的Ethernet. THNet中的交叉开关THSwitch是用15万门的ALTERA FPGA芯片构成，THNet还包括具有DMA引擎的网络适配器THNIA.THNet每一端口可以提供数据传输率为1.056Gbps, 其聚合频宽可达8.4Gbps；采用固定用户缓冲和扩展的主动消息传递等方法，THNet执行用户层的消息传递，旁路操作系统的系统调用，做到零拷贝的消息传递.乒乓测试结果表明：单向消息传递延迟可减少到8μs. THNet软件包括THNIA驱动程序和支持用户层通信的函数库.此文对相关工作进行了简要对比，并说明了该系统的应用情况.
展开▼
