Issue TOC Print full article Print this page Download as PDF E-Mail this URL Flame the author     In this article Introduction Products Reviewed Executive Summary How We Tested Report Card

We tested midrange Fibre Channel switches' throughput and latency, under varying loads (25 percent, 50 percent, 75 percent and 100 percent) and three different frame sizes (60-frame, 1,076-frame and 2,148-frame messages). Our test bed included a Spirent Communications SmartBits 6000B, and SmartBits 600 with SmartFabric software and Spirent's 2-gigabit dual-port Fibre Channel test cards. The SmartBits units were connected directly to the switches under test. We collected test results on a Dell Precision Workstation 410 with 512 MB of RAM and Microsoft Windows Server 2000.

We performed two full-mesh tests--that is, tests in which every port was connected to every other port. The full-mesh throughput test is a killer challenge that measures throughput, high-end latency, frame loss, latency distribution, and gives a latency snapshot. The throughput and frame-loss tests are most important. None of the switches we tested dropped a single frame under any circumstances.

The full-mesh latency test works the same as the full-mesh throughput test, but measures latency. However, because this test really measures buffer latency, the Brocade SilkWorm's extra buffering credits sharply skewed the unit's test results. The Brocade results show that we are clearly measuring frames in the buffer, as the frames have aged in the buffer.

In our port-to-port latency test, we sent 50 frames from one port to another, in pairs only. This tests basic switch latency without any significant load on the switch. Port 0 sends 50 frames to Port 1 and average latency of those frames is measured.

Finally, we created a new test, to measure latency under load and to try to verify Brocade's claims of improved switch performance. We used our SmartBits 600 to measure the port-to-port latency from Port 0 to Port 4 and from Port 8 to Port 12, while our SmartBits 6000B ran a full-mesh test on all the other ports. The result is an uncongested latency measurement while the switch is under load. This is what latency will look like on uncongested links while other links are loaded.