High performance host interfacing for packet-switched networks
Hemant Kanakia · 1992
High performance computer communication between users requires significant improvements over conventional host-to-network interfaces. Current network adapters are either limited in function, wasting key host resources such as the system bus and the processors, or else intelligent but too slow because of complex transport protocols and inadequate internal memory architectures. With networks moving to the gigabit range, these problems will persist in spite of improvements in processor speeds and memory cycle time--unless significant design improvements are achieved. We identified three closely related questions that should be addressed in designing a high performance host interface. The first key question is how to divide transport protocol processing between a network adapter and the main processor. Additional costs of adapters with sufficient intelligence to perform transport protocol has to be balanced against the savings thus made in main processor cycles, host bus and memory bandwidth. The second key question is how much transport-level performance is available with an intelligent adapter. The third key question is how much change in conventional transport protocols is necessary for high performance. We show that intelligent adapters that perform end-to-end checksums, data encryption, and packetization minimize data movement over the host bus and memory, which are critical host resources. Current intelligent adapters are slow and require rethinking of its architecture. We studied three different strategies for increasing the performance of intelligent adapters: pipelined processing, prediction-based header processing, and optimizing latency for packets with small amounts of data. We integrated these three strategies in proposing a Network Adapter Board (NAB) architecture. A prototype for NAB was built for the VMP system, a high-performance multiprocessor workstation developed at Stanford University. The performance for this prototype shows an order of magnitude higher throughput for large data transfer and a third lower latency for small data transfer. Two conclusions, supported by our work, stand out. State machines of current transport protocols do not have to change for high performance. The only required change is the streamlining of transport protocols to facilitate techniques such as pipelined processing, predictive header processing, and optimization of latency for small packets.