/

Reviews /

Backbone switch review: How we did it

Network World Fusion, 10/19/98

We stressed the switches' Gigabit and Fast Ethernet performance (with the exception of the Packet Engines PowerRail 5200, for which the vendor failed to provide 10/100M cards). We tested forwarding rate, latency, congestion and head-of-line (HOL) blocking. All tests, except the HOL blocking test, were run at both Layer 2 and Layer 3.

In Layer 2 tests, the switch is bridging traffic by verifying the media access control cyclic redundancy check (CRC) for each input packet, looking up the MAC layer destination address in its switch table, filtering based on any configured packet attributes, and forwarding the packet without alteration over the appropriate output port.

With Layer 3 switching, each switch routes packets based on their IP network layer address. Traffic from the source is MAC-addressed to the switch, which performs a route table lookup and then forwards the packet MAC-addressed to the next hop or to a directly LAN-attached end system. Layer 3 forwarding involves decrementing the Time to Live field in the IP header, recalculating the header checksum, and recalculating the MAC layer CRC.

We used SmartBits testers from Netcom Systems for all our performance testing, using the SmartApplications and AST test suites for the RFC 1242 and RFC 2285 tests respectively. SmartApplications supports full duplex testing at both Layer 2 and Layer 3 for Ethernet, Fast Ethernet and Gigabit Ethernet, and will be able test true throughput at Layer 2 and Layer 3 as specified in RFC 1242.

(Currently, SmartBits' Layer 2 Gigabit throughput tests are confirmed as accurate, but we were unable to gather reliable results when attempting to run Layer 3 throughput test on the Gigabit ports using the SmartApplications Version 2.05 Beta 3 suite, which was released in August 1998. Netcom Systems planned to correct the problem in hardware, but too late for our testing.)

We measured forwarding rate and packet loss rates for two streams of bidirectional traffic over full duplex Gigabit Ethernet, and 10 streams of bidirectional traffic over full duplex Fast Ethernet. Traffic was not meshed - all the traffic on each input port was forwarded to a single output port. We ran the forwarding rate tests with packet sizes ranging from 64 bytes to 1,518 bytes under loading conditions ranging from 60% to 100%.

We tested latency using a single stream of traffic switched between two different modules in the chassis. Because multilayer switching is necessarily store-and-forward, all latency results are presented in last in, first out format as specified in RFC 1242. Our tests encompassed packet lengths of 64 bytes to 1,518 bytes and loading rates varying from 1% to 98% for both Fast and Gigabit Ethernet.

In the congestion tests, we sent unidirectional traffic from three input ports to a single output port. The tester forwarded bursts of media rate traffic (minimum legal Inter Frame Gap between frames) to each input port. We used an iterative process to determine the maximum burst size each switch could forward without losing packets. We ran the test over a range of packet sizes and for both Fast and Gigabit Ethernet.

A test from the AST suite of Layer 2 benchmarks was run to determine whether the switch exhibits HOL blocking. In this test, one transmit port of the tester delivered a stream of media-rate traffic (100% offered load) destined for a single output port of a switch. A second tester transmit port was programmed with MAC addresses that forced the switch to divide its media rate load equally between the first output port and a second output port. Therefore, one output port was congested by an aggregate load of 150% of media rate, while the second output port was uncongested with a 50% of media rate load. As in previous tests, packet sizes ranged from 64 to 1,518 bytes in length.