Final report for the high performance commodity interconnects for clustered scientific and engineering computing. Part 1 of 3

For a more detailed look, Netpipe was used to provide the signature graphs of the NDIS services over the Giganet VIA and Packet Engines Gigabit Ethernet hardware. Two Hamachi, a second generation Gigabit Ethernet NIC, were installed in place of the first generation cards. From figure 23 the Gigabit Ethernet (GigE) had significant less bandwidth performance although the theoretical line speed was equal. This means that buffering and device tuning would be necessary for such a gateway to function effectively. To examine the TCP/IP transport stack delay, the Giganet VIA latency and the Netpipe TCP stream tests were used. The following assumptions were made during the analysis: (1) The latency introduced by address translation between VIA and TCP/IP is assumed to be negligible compared to the TCP/IP processing. For a real gateway, the lookup would be done via a hash table and set up a priority. (2) Transferring the data from the VIA to the TCP/IP stack would be done through a DMA copy from the VIA NIC to user memory, followed by a memory copy into the TCP/IP stack. The DMA transfer time is assumed to be negligible compared to the memory copy. The latency of the memory copy is included as part of the TCP/IP processing time. The VIA latency test provides a baseline of the inbound VIA. Both the Giganet and the Netpipe application were using the same hardware setup. The time difference between the two provides an insight as to the latency added by the TCP/IP processing. The graph shows the time difference between polling and non-polling VIA latency versus the TCP/IP processing