From The Lab Of Neal Nelson & Associates (1 (312) 755-1000)
Gigabit Ethernet creates a substantial hurdle for organizations that want to test and evaluate this technology: The new devices are so fast that existing tests would not even begin to stress these systems. A new generation of tests is required to match the new generation of networking devices.

To resolve this testing problem, Neal Nelson & Associates (1 (312) 755-1000) in Chicago developed a multinode test bed consisting of 80 Pentium 200/266 computers with 240 10/100 PCI LAN adapters. The test bed operates under an automated control system. It can generate transaction activity of 10,000 users with sustained data rates of more than 10 Gbps using real application programs.

Gigabit Test Bed

To measure the performance of the gigabit devices, we configured 48 of the 266-MHz machines into a single 48-node Linux cluster. The gigabit switches were configured as two groups of 48 ports with a gigabit link connecting them. Tests were run using a combination of ftp, telnet and Web-browser traffic at varying user load levels. A complete description of our test setup as well as a detailed description of the traffic types can be found online.

Results

Our test design provided for different types of traffic and user load levels. We found that the telnet and Web tests reported no degradation regardless of the load level we tested.

The ftp test measured degradation as the user load levels increased. As the test scripts saturated the resources within the gigabit devices, performance degradation occurred. The charts at left and above present data from two of the ftp test configurations.

Our first test, a baseline test at Layer 2, showed that the test bed was capable of sustaining more than 4 Gbps. The highest throughputs were recorded by the products from Extreme and Bay Networks. Even when traffic was confined within the devices on each side and the gigabit link was not active, the different devices delivered varied levels of maximum throughput (2.6 to 4.3 Gbps).

The second test was configured to pass all data through the gigabit link. It was apparent in this test that overall throughput was greatly reduced, compared to the baseline test. The maximum throughput is just under 2 Gbps (1 gigabit in each direction). The throughput spread (1.1 to 1.9 Gbps) shows that some devices filled the gigabit link between the two sides of the test bed more effectively than others.

Our third test measured Layer 3 performance across the gigabit link. Each machine was assigned its own subnet such that all packets in this test had to be routed. Here, throughput was generally slightly lower than the throughput in the Layer 2 test. These results, which you can find online, tend to support gigabit vendors' claims that Layer 3 switching is performed at near wire speed.

--Neal Nelson