Measuring Latency in High-Speed, Low-Latency Networks
Resolution, precision, and accuracy–three words that get thrown around a lot, often interchangeably. That’s not surprising, considering that some dictionaries actually list one or more as synonyms for the other.
But these three words actually mean three different things and that distinction becomes very important when testing devices and networks for latency, particularly in switch fabrics designed for data centers and cloud computing.
... the difference is on how reliable and repeatable a measurement is.
Resolution and precision both refer to the smallest unit measurable by an instrument, the difference is on how reliable and repeatable a measurement is. Neither refers to whether the measurement is correct. That’s what accuracy is all about—correctness.
- Resolution: The smallest unit to which an instrument can be read
- Precision: The smallest unit to which an instrument can be read reliably and repeatably
- Accuracy: The condition or quality of being correct
A common illustration of the difference between resolution and precision compares an analog stopwatch to a digital stopwatch. The analog stopwatch may have ten tick marks between seconds, indicating a resolution of one tenth of a second, whereas the digital stopwatch may have two digits after the decimal, indicating a resolution of one hundredth of a second. However, human response isn’t that fast. It takes about a tenth of a second from a stimulus to a button press. So despite the fact that the digital stopwatch has a resolution of one hundredth of a second, its precision (repeatability) when operated by a human is the same as the analog watch, one tenth of a second.
However, neither resolution nor precision will tell you whether the stopwatch is accurate, meaning that when the stopwatch marks off ten seconds, whether exactly ten seconds has passed. Accuracy in an instrument is usually assured by calibrating it to a known value. In electronic instruments that measure sub-second values, accuracy depends on the resonant frequency of crystals, which can measure millions or billions of changes per second.
Because of the importance of sub-microsecond latency measurements in high-speed environments, Spirent engineers designed Spirent TestCenter to measure time intervals with the highest resolution, precision, and accuracy in the industry. The test modules have a resolution of 10 ns at 10 Gbps and 2.5 ns at 40 Gbps and 100 Gbps. To yield the best results when measuring latency between interfaces of different speeds, the test signature has a resolution of 2.5 ns on all modules.
The shortest frame insertion time is 51.2 ns at 10 Gbps and 20.48 ns at 40 Gbps, so a timestamp resolution of 10 ns is adequate. However, the shortest frame insertion time is 5.12 ns at 100 Gbps, so the 2.5 ns resolution of Spirent TestCenter 100 Gbps test modules is required to accurately measure latency at that speed.
Scalability tests for data centers
require high port counts, which involve
multiple test chassis.
Scalability tests for data centers require high port counts, which involve multiple test chassis. To achieve the accuracy required for high-speed ports at such granular precision, Spirent TestCenter calibrates for synchronization cable length between chassis, port location with each chassis, and heat, as temperature affects the resonant frequency of a crystal. Spirent TestCenter’s ability to display negative latency allows its users to validate calibration and optionally adjust to near zero nanoseconds when using Direct Attached Copper cables vs. fibers with optics.
Recently Gnodal, known for their ultra-low-latency switches, looked to Spirent to showcase the performance of the GS-Series switches, including its latest GS series 40 GbE switch. Gnodal conducted low-latency performance testing of the GS4008 and GS0072 High Speed Ethernet switches individually and also as part of a two-tier data center fabric.
The precision and accuracy of Spirent TestCenter latency measurements become of paramount importance in this test, which measured 40 GbE performance in the low hundreds of nanosecondslatency.
As illustrated by the Gnodal test, modern switch and router fabrics span multiple chassis and typically expanded to build systems supporting hundreds or thousands of ports. These systems must be tested to assess system performance and verify reliability when deployed in large mission-critical environments supporting millions of user application flows. The density and mesh-pattern traffic-generation capabilities of Spirent TestCenter can not only fully-load modern, multi-terabit backplanes and fabrics, but can measure port-to-port and end-to-end latency across a full-mesh network with industry-leading accuracy and precision.
For more information on the importance of accuracy and precision in latency measurements, download the white paper “The Role of Latency and Jitter in Network Performance Assessment.”