HP ProCurve Routing Switch 9308M / 9304M Reviewer’s Guide
latencies should be reported as LIFO. At any rate, latency figures should be labeled as to whether they are LIFO or FIFO.
Occasionally store-and-forward routing switch latency will be reported on a FIFO basis. To convert FIFO to LIFO results, the time associated with the packet itself must be subtracted out. For 10Mbps Ethernet frames the following formula can be used:
LIFO = FIFO - (packet length (bytes) * .8 µsec)
For 100Mbps Ethernet the formula is:
LIFO = FIFO - (packet-length (bytes) * .08 µsec)
For Gigabit Ethernet the formula is:
LIFO = FIFO - (packet-length (bytes) * .008 µsec)
Latency is usually measured with the NetCom SmartBits or Ixia 1600 testers, special-purpose test units that can measure latency down to the nearest bit time for the technology under test.
In normal testing, latency should be measured for a variety of packet sizes and all ports should be tested since latency, being sensitive to the architecture of the switch, can vary from port to port.
As show by the figures in Appendix A, latency for the HP ProCurve Routing Switch 9308M and 9304M is excellent at around 7-17 µsec for packets forwarded on the same module and 10.5-21 µsec for packets that flow across the backplane. The difference in the ‘on module’ latency versus the ‘across the backplane’ latency is the time it takes to actually traverse the backplane and get stored on the target output module. Either way, these latencies are outstanding (very low). And these latencies are with routing active. The architecture of the HP ProCurve routing switches lead to the same packet latencies whether the routing switch is routing IP or IPX or Layer 2 switching the packets because these functions are done in the module ASICs. Multicast packets are handled in hardware incurring the same latency as unicast packets.
For those packets handled by the processor, such as AppleTalk packets or broadcast packets, the latency will run about 190-200 µsec, still a very respectable figure, particularly when compared to the much longer latencies usually seen for all packets in the traditional processor-based router.
