1 draft-duffield-ippm-burst-loss-metrics-01.txt Nick Duffield, Al Morton, AT&T Joel Sommers, Colgate University IETF 76, Hiroshima, Japan 11/10/2009

2 Agenda  History of the draft  One page summary of draft  Mailing list comments and discussion  Related activity  Conclusions

3 History of draft-duffield-ippm-burst-loss-metrics  Aim: standardize measurement of loss episodes [SBDR08]  Initial presentations IETF 72, 73  -00 individual draft published prior to IETF 74  IPR disclosures for -00 draft completed April 2009  -01 draft published July 2009  Open question: should draft be adopted as WG item?  Some comments and questions on draft on the IPPM mailing list  Thanks for comments; more please!

4 A one page summary of the draft  Fact: packets in a flow are not generally loss independently  Motivation: metrics of temporal structure of packet loss  Target use: SLAs, application requirements (e.g VoIP)  Object of study: loss episodes (of consecutively loss packets)  Metrics: average duration and frequency of loss episodes  Probing: bi-packet probes, sent as discrete Poisson stream  Analysis: metrics depend only on frequencies probe outcomes  4 possible outcomes (0,0), (0,1), (1,1), (1,0) where 1 = lost, 0 = not lost  Summary: extension of RFC 2680 to case of correlated loss XXXX X XX X Frequent small glitches vs. local burst (at same average loss rate) (0,0)(0,1)(1,1)(0,0)

5 Mailing list comments and discussion  Should metrics be loss episodes average or general burstiness?  What is the relation to Gilbert model?  Should metrics be time based or count based?  Need for clarification of role of selection function

6 Metrics: Loss episode averages or burstiness? Structure of loss episodes is more complex that average length  Multipacket statistics? (Prob[episode has n packets], n = 1,2,3,…)  Correlations between lengths of episodes, gaps between episodes? Questions/Issues:  What is added utility of multipacket statistics over averages?  Metric statistical accuracy decreases with number of packets n Authors’ Recommendation  Retain only loss episode averages (simple extension of RFC 2680)  Defer multipacket loss statistics as separate WG item if interest NB: averaging metrics do not need to sample full loss episodes

7 What is relation to Gilbert model?  In parametric terms, the Gilbert model is more complex  Gilbert model has 4 parameters: Good/Bad state lifetimes/loss rate  Two independent loss episode metrics (average duration, frequency)  Metrics are purely empirical, interpreted independent of model  Metrics do not aim to estimate parameters of any model  Authors’ Recommendation: expand draft with applicability section

8 Time based or count based episode metrics?  Some well known burstiness metric are based on packet counts  IDC: index of dispersion on counts  Loss episode metrics based on time (average duration etc)  Easier to compare directly with application requirements  Probe rate and traffic rate generally different  Authors’ Recommendation:  Retain time-basis

9 What is the role of selection function?  Selection function is a general formulation of a way to specify which packet are used for probing (see RFC 3393)  Examples:  Specifying how discrete Poisson b-packet probes are to be selected  Potential use to specify selection mechanism for background traffic to be co-opted as probes.  Authors’ Recommendation:  Expand explanation of selection function in draft.

10 Related Activity  ITU Activity  contributed to ITU-T SG 12 Question 17 on packet performance.  independent implementation of same loss episode metrics Special case: unsampled counts of 4 bi-packet outcomes

11 Authors’ Conclusions  Mailing list discussion has been helpful and constructive; thanks!  Points raised appear to request clarifications and elaboration, rather than raising fundamental objections to metrics or methods  Authors will update accordingly in next draft version