A One-way Packet Loss Metric for IPPM
obsoleted by
- rfc-7680 — A One-Way Loss Metric for IP Performance Metrics (IPPM)
Extracted elements (24)
A corrupted packet is counted as lost for consistency: if the IP header is corrupted, the source or destination addresses cannot be verified, making it impossible to confirm the packet corresponds to a given sent test packet. Counting some corrupted packets as received and others as lost would be inconsistent.
One-way loss is measured instead of round-trip loss because Internet paths are often asymmetric (different routers for forward and reverse), asymmetric queueing can create radically different performance characteristics even on symmetric paths, and application performance may depend primarily on one direction. QoS provisioning also often differs per direction.
Poisson sampling is used to schedule test packets to avoid self-synchronization effects and to obtain a statistically unbiased sample. The Framework document provides methods to compute and verify the pseudo-random Poisson process using the Anderson-Darling test.
The loss threshold ('reasonable period of time') is intentionally vague, defined as a value Th so large that any value in [Th-delta, Th+delta] is an equivalent threshold, where delta encompasses all clock synchronization error. A specific non-huge timeout can be approximated by measuring one-way delay and computing the fraction exceeding that value.
Transport connection loss rates (e.g., TCP) do not reflect unbiased samples because TCP transmits in bursts (inducing additional loss) and adapts its rate to minimize loss, making IPPM Poisson-sampled measurements necessary for unbiased network characterization.
For healthy Internet paths operating below 1% loss, discriminating between various fractions of 1% over one-minute periods may require several hundred samples per minute, resulting in larger lambda values than ordinarily desired.
If a packet is fragmented and reassembly does not occur for any reason, the packet is deemed lost. If duplicated packets arrive, the packet is counted as received.
Out-of-order arrival of test packets must be correctly handled in the Poisson stream methodology: a packet sent later (TS[i+1]) may arrive at Dst before a packet sent earlier (TS[i]).
The 255-second theoretical upper bound on IP packet lifetimes (RFC 791) could be used as a simple upper bound for distinguishing loss from very large delay, but good engineering and understanding of packet lifetimes is needed in practice.
Measurement methodologies SHOULD include appropriate techniques to reduce the probability that measurement traffic can be distinguished from normal traffic, to prevent routers from giving it different priority.
Measurement parameters MUST be carefully selected so that measurements inject trivial amounts of additional traffic into the networks they measure, to avoid skewing results or causing congestion and denial of service.
The calibration and context for the underlying singletons MUST be reported along with the stream metric.
The calibration and context in which the metric is measured MUST be carefully considered and SHOULD always be reported along with metric results, including Type-P, loss threshold, instrument calibration, and path traversed.
The degree of synchronization between Src and Dst clocks MUST be reported. If possible, the possibility that a test packet arriving at Dst is reported as lost due to resource exhaustion SHOULD be reported.
The exact Type-P used to make measurements MUST be accurately reported, as the metric value may depend on protocol (UDP/TCP), port number, size, or special treatment (e.g., IP precedence or RSVP).
The path traversed by the packet SHOULD be reported if possible, acknowledging that knowing the precise path is generally impractical but partial information (e.g., initial link choice) is still valuable context.
The threshold or methodology used to distinguish between a large finite delay and loss MUST be reported.
The privacy concerns of network measurement are limited by the active measurement approach: unlike passive measurements, there can be no release of existing user data since only injected test packets are observed.
Four clock uncertainty concepts are defined for measurement methodology: synchronization (agreement between two clocks), accuracy (agreement with UTC), resolution (precision of a clock), and skew (change of accuracy or synchronization over time).
Type-P-One-way-Packet-Loss-Average is the arithmetic mean of all L values in a Type-P-One-way-Packet-Loss-Poisson-Stream. It is undefined if the sample is empty.
Type-P-One-way-Packet-Loss is 0 when Src sent the first bit of a Type-P packet at wire-time T and Dst received it; it is 1 when Dst did not receive the packet. It is 0 exactly when Type-P-One-way-Delay is finite, and 1 exactly when Type-P-One-way-Delay is undefined.
Type-P-One-way-Packet-Loss is a singleton metric with parameters Src (IP address), Dst (IP address), and T (a time). Its value is 0 (packet received) or 1 (packet lost).
Type-P-One-way-Packet-Loss-Poisson-Stream is a sample metric with parameters Src, Dst, T0 (start time), Tf (end time), and lambda (rate in reciprocal seconds). It produces a sequence of <time, loss> pairs where times are drawn from a pseudo-random Poisson process between T0 and Tf.
Active measurements inject packets into the network and can skew results or cause congestion and denial of service if too much traffic is injected. Additionally, routers recognizing measurement traffic and treating it differently, or attackers injecting artificial measurement traffic to artificially lower observed loss rates, are threats to measurement integrity.