Full Mesh Throughput
click to enlarge
Performance testing, a key part of our review because of the importance of the traffic the switches carry, gave us a wild ride. Nervous vendors ushered their switches into Network Computing's Green Bay, Wis., Real-World Labs® like parents taking their children to kindergarten for the first time. We conducted full-mesh testing, in which every port sends frames to every other port; frame-loss testing; and several latency tests, including full-mesh, standard industry latency and our own latency-under-load test. Only the frame-loss tests came out exactly as expected: None of the switches dropped a single frame, even when we accidentally misconfigured a test.
None of these products can be characterized as a bad performer; however, results did vary. QLogic's SANbox 2 showed consistent numbers across the board. But Brocade SilkWorm 3800 gave us some squirrelly test results. At 100 percent load on the full-mesh latency test, for example, the Brocade switch's latency ranged from seven to 85 times that of the second-place competitor. Brocade attributed this anomaly to the SilkWorm 3800's extra buffering algorithms, which threw the results off because our test measures buffer latency only. (Brocade also does not have a single integrated circuit, which the McData and QLogic products have, nor is its buffer RAM integrated on the IC.)
McData's Sphereon 4500 showed some good latency numbers, especially in the full-mesh latency tests. Curiously, that switch also showed the worst numbers in the industry-standard port-to-port tests but very good results in latency-under-load tests and full-mesh latency tests.