Loss Episode Metrics for IP Performance Metrics (IPPM)
Extracted elements (18)
Bi-packet probing estimates average loss episode duration without directly measuring complete episodes; the information is encoded in the relative frequencies of the four Loss Pair outcomes. This is more efficient than the direct episode-reporting approach of RFC 3611 and does not require observing every lost packet in an episode.
Geometric inter-pair spacing is chosen so that a general geometric-stream sample with launch probability q < 1 independently sub-samples each Loss Pair of the idealized q=1 stream, ensuring that normalized Loss-Pair-Counts are on average identical to the idealized sample and yielding unbiased estimators for the loss episode metrics.
Loss episode metrics are defined as functions of Loss-Pair-Counts rather than sequential loss patterns because geometric-stream samples do not report all lost packets within each episode; the count-based formulas remain valid estimators without requiring complete episode observation.
The document deliberately moves from a single loss metric (average packet loss ratio per RFC 2680) to a pair of metrics — loss episode frequency and average loss episode duration — as the minimal generalization that captures bursty loss structure in a model-independent, empirical manner directly comparable to application tolerance for loss patterns.
The loss episode metrics of this memo subsume the Gilbert model: the two Gilbert transition probabilities P(g|b) = d/m and P(b|g) = (d/m)/(1/r - 1) can be recovered algebraically from the loss ratio r and episode duration m, showing the metrics are sufficient to parameterize the simplest bursty-loss Markov model without requiring Markovian assumptions during measurement.
Out-of-order packet arrival must be handled in any active bi-packet loss measurement implementation; implementations should follow the guidance in Section 3.6 of RFC 2680 for this case.
The bi-packet loss metrics are derived metrics per RFC 2330 Section 6.1, built on the singleton loss metric of RFC 2680 Section 2; the methodologies, error sources, and reporting conventions of RFC 2680 apply to each individual packet within every probe pair.
Measurement methodologies SHOULD include appropriate techniques to reduce the probability that measurement traffic can be distinguished from normal traffic, to prevent routers from treating probes differently and to prevent attackers from artificially lowering the observed loss rate by injecting fake probe traffic.
The measurement parameters MUST be carefully selected so that the measurements inject only trivial amounts of additional traffic into the networks they measure; injecting too much traffic can skew the results and, in extreme cases, cause congestion and denial of service.
The privacy concerns of these network measurements are limited because they involve only active probe packets injected by the measurement system and no release of user data.
Bi-Packet-Loss-Episode-Duration-Number estimates average loss episode length in packets: (2*N(1,1) + N(0,1) + N(1,0)) / (N(0,1) + N(1,0)) when loss transitions exist; 0 if no packets are lost; undefined if all probe packets are lost.
Bi-Packet-Loss-Episode-Frequency-Number estimates how often loss episodes begin per probe launch, computed as Bi-Packet-Loss-Ratio divided by Bi-Packet-Loss-Episode-Duration-Number, with special cases: 0 if no packets are lost, and 1 if all probe packets are lost.
Bi-Packet-Loss-Ratio is defined as (N(1,0) + N(1,1))/n and is formally equivalent to the average single-packet loss ratio Type-P-One-way-Packet-Loss-Average from RFC 2680, since it averages single-packet losses.
Loss-Pair-Counts N(i,j) are the absolute frequencies of the four Loss Pair outcome types — (0,0), (0,1), (1,0), (1,1) — in a sample of n Loss Pairs. They are the fundamental building block for all derived loss episode proto-metrics and final metrics.
Type-P-One-way-Bi-Packet-Loss-Geometric-Stream-Episode-Frequency is the Bi-Packet-Loss-Episode-Frequency-Number divided by d, yielding the estimated average rate per second at which loss episodes start (or finish). Multiplying by d recovers the frequency relative to probe launches.
Type-P-One-way-Bi-Packet-Loss is the fundamental singleton metric, defined as a Loss Pair (l1, l2) where each of l1 and l2 is binary (0 = received, 1 = lost), capturing the four possible outcomes for a pair of packets sent from Src to Dst at wire-times T1 and T2.
Type-P-One-way-Bi-Packet-Loss-Stream is a sample of n Loss Pairs from n probe pairs sent at times (T11,T12), ..., (Tn1,Tn2). It is general enough to subsample a pre-existing measurement stream; bi-packet probes need not be consecutive within such a stream.
Active measurements face two threats: injecting too many probe packets risks skewing results or causing denial of service; and an attacker injecting artificial traffic accepted as legitimate probes will artificially lower the observed loss rate. Authentication techniques such as digital signatures may be used to guard against injected traffic attacks.