scope:

This Recommendation specifies metrics of latency under simultaneous traffic load and defines methods of measurement to increase the specificity and repeatability of metric assessment.

Key points defining this work are:

• Multiple metrics are required to assist with interpreting the results, such as the minimum delay, and one or more metrics of delay when the path is operating at maximum capacity.

• It is important to distinguish the network latency from the contribution to latency in end-hosts, since the latter varies with hosts and falls under different administrations. This Recommendation focuses on measuring the network latency.

• All metrics need to be carefully defined. Latency is a common term, and round-trip time or round-trip packet delay are more exact. Statistical measures will be used to summarize the many individual measurements of packet delay collected during a measurement interval.

• The methods are primarily intended for access paths to the Internet, possibly extending to one or more Internet gateways. Multiple transport protocols will be considered.

• Mobile networks are included in the scope, and with them the possibility of variable capacity over time.

• Bimodal capacity is characterized by the presence of two-time ranges where capacity, observed by repeated measurements, is distinguishably different. Typically, but not exclusively so, the capacity during the initial seconds of transmission is high, followed by a stable, lower capacity for the remainder of the transmission. It is intended to characterize the latency under load for each mode when present. Multimodal paths may also be present, requiring the characterization of each mode. There are at least two audiences for latency under load metrics: technical experts will be able to understand network performance metrics, but typical consumers do not (yet) have an appreciation of their needs on the latency continuum in terms of milliseconds. The results may need to be expressed in categories for consumers (such as red, yellow, green).

Note that technically the latency experienced by a packet queued in a buffer of constant depth depends on the capacity to forward (dequeue) the temporary buffered packets. If the dequeuing capacity is variable, so is the latency experienced by repeated measurement flows. The latency is proportional to the inverse of the dequeue capacity.