Lies, Damned Lies and Latency

Some vendors' data sheets claim lower latency and jitter than the numbers we measured in this test. While these claims are valid, they also should raise questions about how the measurements were obtained.

Test engineers can employ at least two tricks to shave off time from latency and jitter measurements. The first is to measure when offering traffic only between one or more pairs of ports. This pattern, called "non-meshed" in RFC 2285 on Ethernet switch testing, tends to be less stressful than using partially or fully meshed patterns, where all ports exchange frames. Port-pair testing can be like "proving" a car will go very fast, but only if the driver never turns left or right.

A second technique is to characterize and subtract out the propagation time added by cables and fiber or copper transceivers. Cabling can add significant propagation delay, given long enough cables, and certainly should not be charged to the system under test – but it wasn't a factor in this test. We used 1-meter cables, introducing a total of about 10 nanosec propagation time – and that 10 nanosec also happens to be the timestamp resolution of Spirent TestCenter. Put another way, since any latency or jitter measurement is accurate within plus or minus 10 nanosec, cable propagation time was not significant in our tests.

Switch transceivers also add delay and jitter, but subtracting out that time is not valid for end-users. This practice makes sense for hardware and test engineers looking to characterize the performance of a switch fabric in isolation. But end-users buy a switch, not a fabric. In production settings, the latency and jitter measurements that matter are those observed at the front-side ports using stressful meshed traffic patterns.

Return to main test.

Join the Network World communities on Facebook and LinkedIn to comment on topics that are top of mind.
Related:
Must read: 10 new UI features coming to Windows 10