Network World, 8/11/97

We delivered a full load of traffic packets to each 100Base-T switch based on the number of ports in the switch's configuration. That is, we delivered a stream of packets - at the maximum Fast Ethernet rate the switch would accept - to half of the switch's ports at once.

For many switches, the maximum packet stream is 148,800 packet/sec. Some switches, however, can accept a maximum packet stream only at a slightly lower rate - typically in the 145,000- to 148,000-packet/sec range. After we determined each switch's rate of maximum acceptance, we generated load streams using Shomiti Systems, Inc.'s Century Analyzers.

Before conducting any load test through a switch, we first disabled that switch's flow control, if possible. Flow control lets the switch throttle back individual input ports when their rate of packet delivery is too high for the switch to handle and when the result would otherwise be packets dropped or lost. Note that, in some cases, a switch's maximum throughput is slightly higher with flow control enabled.

The traffic streams we delivered were unidirectional and consisted of all 64-byte packets addressed to nodes known to the switch (with their unicast media access control address already learned and entered into the switch's address table).

We delivered our traffic streams onto the switch's odd-numbered ports, with the packets in each stream destined for one of the switch's even-numbered ports. The packets in each stream had to be individually switched to output ports in round-robin fashion. Packets delivered on port 1, for instance, were destined for ports 2, 4, 6, 8, 2, 4, 6 and so on. The test traffic streams were constructed in part with the help of Midnight Networks, Inc.'s Avalanche network-simulation tool.

We also looked for indications of head-of-line blocking. In general, head-of-line blocking refers to degradation of throughput on a particular switch port caused by conditions, such as a port overload, occurring elsewhere in the switch.

To check for head-of-line blocking, we used Hewlett-Packard Co.'s Internet Advisors to deliver a high-volume packet stream via an input port, with traffic destined to two output ports. The stream would direct a fairly high load of, say, 70% of capacity to output port A and a second ''trickle'' packet stream, a load of only about 10% of capacity, to a different output port B.

A second packet stream of about 50% capacity would be delivered, via a different input port, with all of its traffic destined for output port A, to overload the output port. We then used an Internet Advisor to check for head-of-line blocking by observing any throughput degradation on the trickle stream on output port B.