Impact of implementation on XTP throughput performance

An analysis of the throughput performance of a software implementation of the Xpress Transfer Protocol (XTP) is presented. An instruction level simulator is used to measure performance of the XTP software. Instruction paths that make up the send and receive threads of protocol processing as well as checksum processing are timed to provide measurements of the minimum delay in the transfer layer. The sensitivity of the throughput to variations in several implementation related parameter values is examined. It is shown that the dominant parameter determining performance is the per byte processing. Additionally, it is shown that implementation of the checksum calculation in software imposes a significant penalty on throughput. It is also shown that the effect of per packet and per message processing delays increases with reduced per byte processing, and the effect of packet sizes is examined. The order of improvement in these parameters that will be needed to achieve 100Mb/s or higher speeds is studied