A High Performance Network Architecture for the HP Apollo 9000 Series 700 Workstation

With current low-cost, high performance workstations, application-to-application throughput is limited more by host memory bandwidth than by the cost of protocol processing. Conventional network architectures are inefficient in their use of this memory bandwidth, because data is copied several times between the application and the network. As networks speeds increase further, network architectures must be developed that reduce the demands on host memory bandwidth. In this paper we discuss the design of a single copy network architecture, where data is copied directly between the application buffer and the network interface. Protocol processing is performed by the host, and transport layer buffering is provided on the network interface. We describe a prototype implementation, for the HP Apollo Series 700 workstation family, that consists of an FDDI network interface and a modified 4.3BSD TCPIIP protocol stack, and we report some early results that demonstrate twice the throughput of a conventional network architecture, and significantly lower latency.