RFC 8382






Internet Engineering Task Force (IETF)                     D. Hayes, Ed.
Request for Comments: 8382                                     S. Ferlin
Category: Experimental                        Simula Research Laboratory
ISSN: 2070-1721                                                 M. Welzl
                                                               K. Hiorth
                                                      University of Oslo
                                                               June 2018


Shared Bottleneck Detection for Coupled Congestion Control for RTP Media

Abstract



   This document describes a mechanism to detect whether end-to-end data
   flows share a common bottleneck.  This mechanism relies on summary
   statistics that are calculated based on continuous measurements and
   used as input to a grouping algorithm that runs wherever the
   knowledge is needed.

Status of This Memo



   This document is not an Internet Standards Track specification; it is
   published for examination, experimental implementation, and
   evaluation.

   This document defines an Experimental Protocol for the Internet
   community.  This document is a product of the Internet Engineering
   Task Force (IETF).  It represents the consensus of the IETF
   community.  It has received public review and has been approved for
   publication by the Internet Engineering Steering Group (IESG).  Not
   all documents approved by the IESG are candidates for any level of
   Internet Standard; see Section 2 of RFC 7841.

   Information about the current status of this document, any errata,
   and how to provide feedback on it may be obtained at
   https://www.rfc-editor.org/info/rfc8382.















Hayes, et al.                 Experimental                      [Page 1]

RFC 8382                SBD for CCC for RTP Media              June 2018


Copyright Notice



   Copyright (c) 2018 IETF Trust and the persons identified as the
   document authors.  All rights reserved.

   This document is subject to BCP 78 and the IETF Trust's Legal
   Provisions Relating to IETF Documents
   (https://trustee.ietf.org/license-info) in effect on the date of
   publication of this document.  Please review these documents
   carefully, as they describe your rights and restrictions with respect
   to this document.  Code Components extracted from this document must
   include Simplified BSD License text as described in Section 4.e of
   the Trust Legal Provisions and are provided without warranty as
   described in the Simplified BSD License.





































Hayes, et al.                 Experimental                      [Page 2]

RFC 8382                SBD for CCC for RTP Media              June 2018


Table of Contents



   1. Introduction ....................................................4
      1.1. The Basic Mechanism ........................................4
      1.2. The Signals ................................................4
           1.2.1. Packet Loss .........................................4
           1.2.2. Packet Delay ........................................5
           1.2.3. Path Lag ............................................5
   2. Definitions .....................................................6
      2.1. Parameters and Their Effects ...............................7
      2.2. Recommended Parameter Values ...............................8
   3. Mechanism .......................................................9
      3.1. SBD Feedback Requirements .................................10
           3.1.1. Feedback When All the Logic Is Placed at
                  the Sender .........................................10
           3.1.2. Feedback When the Statistics Are Calculated at the
                  Receiver and SBD Is Performed at the Sender ........11
           3.1.3. Feedback When Bottlenecks Can Be Determined
                  at Both Senders and Receivers ......................11
      3.2. Key Metrics and Their Calculation .........................12
           3.2.1. Mean Delay .........................................12
           3.2.2. Skewness Estimate ..................................12
           3.2.3. Variability Estimate ...............................13
           3.2.4. Oscillation Estimate ...............................13
           3.2.5. Packet Loss ........................................14
      3.3. Flow Grouping .............................................14
           3.3.1. Flow-Grouping Algorithm ............................14
           3.3.2. Using the Flow Group Signal ........................18
   4. Enhancements to the Basic SBD Algorithm ........................18
      4.1. Reducing Lag and Improving Responsiveness .................18
           4.1.1. Improving the Response of the Skewness Estimate ....19
           4.1.2. Improving the Response of the Variability
                  Estimate ...........................................20
      4.2. Removing Oscillation Noise ................................21
   5. Measuring OWD ..................................................21
      5.1. Timestamp Resolution ......................................21
      5.2. Clock Skew ................................................22
   6. Expected Feedback from Experiments .............................22
   7. IANA Considerations ............................................22
   8. Security Considerations ........................................22
   9. References .....................................................23
      9.1. Normative References ......................................23
      9.2. Informative References ....................................23
   Acknowledgments ...................................................25
   Authors' Addresses ................................................25






Hayes, et al.                 Experimental                      [Page 3]

RFC 8382                SBD for CCC for RTP Media              June 2018


1.  Introduction



   In the Internet, it is not normally known whether flows (e.g., TCP
   connections or UDP data streams) traverse the same bottlenecks.  Even
   flows that have the same sender and receiver may take different paths
   and may or may not share a bottleneck.  Flows that share a bottleneck
   link usually compete with one another for their share of the
   capacity.  This competition has the potential to increase packet loss
   and delays.  This is especially relevant for interactive applications
   that communicate simultaneously with multiple peers (such as
   multi-party video).  For RTP media applications such as RTCWEB,
   [RTP-COUPLED-CC] describes a scheme that combines the congestion
   controllers of flows in order to honor their priorities and avoid
   unnecessary packet loss as well as delay.  This mechanism relies on
   some form of Shared Bottleneck Detection (SBD); here, a measurement-
   based SBD approach is described.

1.1.  The Basic Mechanism



   The mechanism groups flows that have similar statistical
   characteristics together.  Section 3.3.1 describes a simple method
   for achieving this; however, a major part of this document is
   concerned with collecting suitable statistics for this purpose.

1.2.  The Signals



   The current Internet is unable to explicitly inform endpoints as to
   which flows share bottlenecks, so endpoints need to infer this from
   whatever information is available to them.  The mechanism described
   here currently utilizes packet loss and packet delay but is not
   restricted to these.  As Explicit Congestion Notification (ECN)
   becomes more prevalent, it too will become a valuable base signal
   that can be correlated to detect shared bottlenecks.

1.2.1.  Packet Loss



   Packet loss is often a relatively infrequent indication that a flow
   traverses a bottleneck.  Therefore, on its own it is of limited use
   for SBD; however, it is a valuable supplementary measure when it is
   more prevalent (refer to [RFC7680], Section 2.5 for measuring packet
   loss).










Hayes, et al.                 Experimental                      [Page 4]

RFC 8382                SBD for CCC for RTP Media              June 2018


1.2.2.  Packet Delay



   End-to-end delay measurements include noise from every device along
   the path, in addition to the delay perturbation at the bottleneck
   device.  The noise is often significantly increased if the round-trip
   time is used.  The cleanest signal is obtained by using One-Way Delay
   (OWD) (refer to [RFC7679], Section 3 for a definition of OWD).

   Measuring absolute OWD is difficult, since it requires both the
   sender and receiver clocks to be synchronized.  However, since the
   statistics being collected are relative to the mean OWD, a relative
   OWD measurement is sufficient.  Clock skew is not usually significant
   over the time intervals used by this SBD mechanism (see [RFC6817],
   Appendix A.2 for a discussion on clock skew and OWD measurements).
   However, in circumstances where it is significant, 5.2">Section 5.2
   outlines a way of adjusting the calculations to cater to it.

   Each packet arriving at the bottleneck buffer may experience very
   different queue lengths and, therefore, different waiting times.  A
   single OWD sample does not, therefore, characterize the path well.
   However, multiple OWD measurements do reflect the distribution of
   delays experienced at the bottleneck.

1.2.3.  Path Lag



   Flows that share a common bottleneck may traverse different paths,
   and these paths will often have different base delays.  This makes it
   difficult to correlate changes in delay or loss.  This technique uses
   the long-term shape of the delay distribution as a base for
   comparison to counter this.





















Hayes, et al.                 Experimental                      [Page 5]

RFC 8382                SBD for CCC for RTP Media              June 2018


2.  Definitions



   The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
   "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and
   "OPTIONAL" in this document are to be interpreted as described in
   BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all
   capitals, as shown here.

   Acronyms used in this document:

      OWD - One-Way Delay

      MAD - Mean Absolute Deviation

      SBD - Shared Bottleneck Detection

   Conventions used in this document:

      T            the base time interval over which measurements
                   are made

      N            the number of base time, T, intervals used in some
                   calculations

      M            the number of base time, T, intervals used in some
                   calculations, where M <= N

      sum(...)     summation of terms of the variable in parentheses

      sum_T(...)   summation of all the measurements of the variable in
                   parentheses taken over the interval T

      sum_NT(...)  summation of all measurements taken over the
                   interval N*T

      sum_MT(...)  summation of all measurements taken over the
                   interval M*T

      E_T(...)     the expectation or mean of the measurements of the
                   variable in parentheses over T

      E_N(...)     the expectation or mean of the last N values of the
                   variable in parentheses

      E_M(...)     the expectation or mean of the last M values of the
                   variable in parentheses





Hayes, et al.                 Experimental                      [Page 6]

RFC 8382                SBD for CCC for RTP Media              June 2018


      num_T(...)   the count of measurements of the variable in
                   parentheses taken in the interval T

      num_MT(...)  the count of measurements of the variable in
                   parentheses taken in the interval M*T

      PB           a boolean variable indicating that the particular
                   flow was identified transiting a bottleneck in the
                   previous interval T (i.e., "Previously Bottleneck")

      skew_est     a measure of skewness in an OWD distribution

      skew_base_T  a variable used as an intermediate step in
                   calculating skew_est

      var_est      a measure of variability in OWD measurements

      var_base_T   a variable used as an intermediate step in
                   calculating var_est

      freq_est     a measure of low-frequency oscillation in the OWD
                   measurements

      pkt_loss     a measure of the proportion of packets lost

      p_l, p_f, p_mad, c_s, c_h, p_s, p_d, p_v
                   various thresholds used in the mechanism

      M and F      number of values related to N

2.1.  Parameters and Their Effects



   T         T should be long enough so that there are enough packets
             received during T for a useful estimate of the short-term
             mean OWD and variation statistics.  Making T too large can
             limit the efficacy of freq_est.  It will also increase the
             response time of the mechanism.  Making T too small will
             make the metrics noisier.

   N and M   N should be large enough to provide a stable estimate of
             oscillations in OWD.  Often, M=N is just fine, though
             having M<N may be beneficial in certain circumstances.  M*T
             needs to be long enough to provide stable estimates of
             skewness and MAD.







Hayes, et al.                 Experimental                      [Page 7]

RFC 8382                SBD for CCC for RTP Media              June 2018


   F         F determines the number of intervals over which statistics
             are considered to be equally weighted.  When F=M, recent
             and older measurements are considered equal.  Making F<M
             can increase the responsiveness of the SBD mechanism.  If F
             is too small, statistics will be too noisy.

   c_s       c_s is the threshold in skew_est used for determining
             whether a flow is transiting a bottleneck or not.  Lower
             values of c_s require bottlenecks to be more congested to
             be considered for grouping by the mechanism.  c_s should be
             set within the range of +0.2 to -0.1 -- low enough so that
             lightly loaded paths do not give a false indication.

   p_l       p_l is the threshold in pkt_loss used for determining
             whether a flow is transiting a bottleneck or not.  When
             pkt_loss is high, it becomes a better indicator of
             congestion than skew_est.

   c_h       c_h adds hysteresis to the bottleneck determination.  It
             should be large enough to avoid constant switching in the
             determination but low enough to ensure that grouping is not
             attempted when there is no bottleneck and the delay and
             loss signals cannot be relied upon.

   p_v       p_v determines the sensitivity of freq_est to noise.
             Making it smaller will yield higher but noisier values for
             freq_est.  Making it too large will render it ineffective
             for determining groups.

   p_*       Flows are separated when the
             skew_est|var_est|freq_est|pkt_loss measure is greater than
             p_s|p_mad|p_f|p_d.  Adjusting these is a compromise between
             false grouping of flows that do not share a bottleneck and
             false splitting of flows that do.  Making them larger can
             help if the measures are very noisy, but reducing the noise
             in the statistical measures by adjusting T and N|M may be a
             better solution.

2.2.  Recommended Parameter Values



   [Hayes-LCN14] uses T=350ms and N=50.  The other parameters have been
   tightened to reflect minor enhancements to the algorithm outlined in
   Section 4: c_s=0.1, p_f=p_d=0.1, p_s=0.15, p_mad=0.1, p_v=0.7.  M=30,
   F=20, and c_h=0.3 are additional parameters defined in that document.
   These are values that seem to work well over a wide range of
   practical Internet conditions.





Hayes, et al.                 Experimental                      [Page 8]

RFC 8382                SBD for CCC for RTP Media              June 2018


3.  Mechanism



   The mechanism described in this document is based on the observation
   that when flows traverse a common bottleneck, each flow's
   distribution of packet delay measurements has similar shape
   characteristics.  These shape characteristics are described using
   three key summary statistics --

   1.  variability estimate (var_est; see Section 3.2.3)

   2.  skewness estimate (skew_est; see Section 3.2.2)

   3.  oscillation estimate (freq_est; see Section 3.2.4)

   -- with packet loss (pkt_loss; see Section 3.2.5) used as a
   supplementary statistic.

   Summary statistics help to address both the noise and the path lag
   problems by describing the general shape over a relatively long
   period of time.  Each summary statistic portrays a "view" of the
   bottleneck link characteristics, and when used together, they provide
   a robust discrimination for grouping flows.  An RTP media device may
   be both a sender and a receiver.  SBD can be performed at either a
   sender or a receiver, or both.

   In Figure 1, there are two possible locations for shared bottleneck
   detection: the sender side and the receiver side.

                                  +----+
                                  | H2 |
                                  +----+
                                     |
                                     | L2
                                     |
                         +----+  L1  |  L3  +----+
                         | H1 |------|------| H3 |
                         +----+             +----+

   A network with three hosts (H1, H2, H3) and three links (L1, L2, L3)

                                 Figure 1

   1.  Sender side: Consider a situation where host H1 sends media
       streams to hosts H2 and H3, and L1 is a shared bottleneck.  H2
       and H3 measure the OWD and packet loss and periodically send
       either this raw data or the calculated summary statistics to H1
       every T.  H1, having this knowledge, can determine the shared
       bottleneck and accordingly control the send rates.



Hayes, et al.                 Experimental                      [Page 9]

RFC 8382                SBD for CCC for RTP Media              June 2018


   2.  Receiver side: Consider that H2 is also sending media to H3, and
       L3 is a shared bottleneck.  If H3 sends summary statistics to H1
       and H2, neither H1 nor H2 alone obtains enough knowledge to
       detect this shared bottleneck; H3 can, however, determine it by
       combining the summary statistics related to H1 and H2,
       respectively.

3.1.  SBD Feedback Requirements



   There are three possible scenarios, each with different feedback
   requirements:

   1.  Both summary statistic calculations and SBD are performed at
       senders only.  When sender-based congestion control is
       implemented, this method is RECOMMENDED.

   2.  Summary statistics are calculated on the receivers, and SBD is
       performed at the senders.

   3.  Summary statistic calculations are performed on receivers, and
       SBD is performed at both senders and receivers (beyond the scope
       of this document, but allows cooperative detection of
       bottlenecks).

   All three possibilities are discussed for completeness in this
   document; however, it is expected that feedback will take the form of
   scenario 1 and operate in conjunction with sender-based congestion
   control mechanisms.

3.1.1.  Feedback When All the Logic Is Placed at the Sender



   Having the sender calculate the summary statistics and determine the
   shared bottlenecks based on them has the advantage of placing most of
   the functionality in one place -- the sender.

   For every packet, the sender requires accurate relative OWD
   measurements of adequate precision, along with an indication of lost
   packets (or the proportion of packets lost over an interval).  A
   method to provide such measurement data with the RTP Control Protocol
   (RTCP) is described in [RTCP-CC-FEEDBACK].

   Sums, var_base_T, and skew_base_T are calculated incrementally as
   relative OWD measurements are determined from the feedback messages.
   When the mechanism has received sufficient measurements to cover the
   base time interval T for all flows, the summary statistics (see
   Section 3.2) are calculated for that T interval and flows are grouped
   (see Section 3.3.1).  The exact timing of these calculations will
   depend on the frequency of the feedback message.



Hayes, et al.                 Experimental                     [Page 10]

RFC 8382                SBD for CCC for RTP Media              June 2018


3.1.2.  Feedback When the Statistics Are Calculated at the Receiver and
        SBD Is Performed at the Sender



   This scenario minimizes feedback but requires receivers to send
   selected summary statistics at an agreed-upon regular interval.  We
   envisage the following exchange of information to initialize the
   system:

   o  An initialization message from the sender to the receiver will
      contain the following information:

      *  A list of which key metrics should be collected and relayed
         back to the sender out of a possibly extensible set (pkt_loss,
         var_est, skew_est, and freq_est).  The grouping algorithm
         described in this document requires all four of these metrics,
         and receivers MUST be able to provide them, but future
         algorithms may be able to exploit other metrics (e.g., metrics
         based on explicit network signals).

      *  The values of T, N, and M, and the necessary resolution and
         precision of the relayed statistics.

   o  A response message from the receiver acknowledges this message
      with a list of key metrics it supports (subset of the sender's
      list) and is able to relay back to the sender.

   This initialization exchange may be repeated to finalize the set of
   metrics that will be used.  All agreed-upon metrics need to be
   supported by all receivers.  It is also recommended that an
   identifier for the SBD algorithm version be included in the
   initialization message from the sender, so that potential advances in
   SBD technology can be easily deployed.  For reference, the mechanism
   outlined in this document has the identifier "SBD=01".

   After initialization, the agreed-upon summary statistics are fed back
   to the sender (nominally every T).

3.1.3.  Feedback When Bottlenecks Can Be Determined at Both Senders and
        Receivers



   This type of mechanism is currently beyond the scope of the SBD
   algorithm described in this document.  It is mentioned here to ensure
   that sender/receiver cooperative shared bottleneck determination
   mechanisms that are more advanced remain possible in the future.

   It is envisaged that such a mechanism would be initialized in a
   manner similar to that described in Section 3.1.2.




Hayes, et al.                 Experimental                     [Page 11]

RFC 8382                SBD for CCC for RTP Media              June 2018


   After initialization, both summary statistics and shared bottleneck
   determinations should be exchanged, nominally every T.

3.2.  Key Metrics and Their Calculation



   Measurements are calculated over a base interval (T) and summarized
   over N or M such intervals.  All summary statistics can be calculated
   incrementally.

3.2.1.  Mean Delay



   The mean delay is not a useful signal for comparisons between flows,
   since flows may traverse quite different paths and clocks will not
   necessarily be synchronized.  However, it is a base measure for the
   three summary statistics.  The mean delay, E_T(OWD), is the average
   OWD measured over T.

   To facilitate the other calculations, the last N E_T(OWD) values will
   need to be stored in a cyclic buffer along with the moving average of
   E_T(OWD):

      mean_delay = E_M(E_T(OWD)) = sum_M(E_T(OWD)) / M

   where M <= N.  Setting M to be less than N allows the mechanism to be
   more responsive to changes, but potentially at the expense of a
   higher error rate (see Section 4.1 for a discussion on improving the
   responsiveness of the mechanism).

3.2.2.  Skewness Estimate



   Skewness is difficult to calculate efficiently and accurately.
   Ideally, it should be calculated over the entire period (M*T) from
   the mean OWD over that period.  However, this would require storing
   every delay measurement over the period.  Instead, an estimate is
   made over M*T based on a calculation every T using the previous T's
   calculation of mean_delay.

   The base for the skewness calculation is estimated using a counter
   initialized every T.  It increments for OWD samples below the mean
   and decrements for OWD above the mean.  So, for each OWD sample:

      if (OWD < mean_delay) skew_base_T++

      if (OWD > mean_delay) skew_base_T--







Hayes, et al.                 Experimental                     [Page 12]

RFC 8382                SBD for CCC for RTP Media              June 2018


   mean_delay does not include the mean of the current T interval to
   enable it to be calculated iteratively.

   skew_est = sum_MT(skew_base_T) / num_MT(OWD)

      where skew_est is a number between -1 and 1.

   Note: Care must be taken when implementing the comparisons to ensure
   that rounding does not bias skew_est.  It is important that the mean
   is calculated with a higher precision than the samples.

3.2.3.  Variability Estimate



   Mean Absolute Deviation (MAD) is a robust variability measure that
   copes well with different send rates.  It can be implemented in an
   online manner as follows:

      var_base_T = sum_T(|OWD - E_T(OWD)|)

         where

            |x| is the absolute value of x

            E_T(OWD) is the mean OWD calculated in the previous T

      var_est = MAD_MT = sum_MT(var_base_T) / num_MT(OWD)

3.2.4.  Oscillation Estimate



   An estimate of the low-frequency oscillation of the delay signal is
   calculated by counting and normalizing the significant mean,
   E_T(OWD), crossings of mean_delay:

      freq_est = number_of_crossings / N

         where we define a significant mean crossing as a crossing that
         extends p_v * var_est from mean_delay.  In our experiments, we
         have found that p_v = 0.7 is a good value.













Hayes, et al.                 Experimental                     [Page 13]

RFC 8382                SBD for CCC for RTP Media              June 2018


   freq_est is a number between 0 and 1.  freq_est can be approximated
   incrementally as follows:

   o  With each new calculation of E_T(OWD), a decision is made as to
      whether this value of E_T(OWD) significantly crosses the current
      long-term mean, mean_delay, with respect to the previous
      significant mean crossing.

   o  A cyclic buffer, last_N_crossings, records a 1 if there is a
      significant mean crossing; otherwise, it records a 0.

   o  The counter, number_of_crossings, is incremented when there is a
      significant mean crossing and decremented when a non-zero value is
      removed from the last_N_crossings.

   This approximation of freq_est was not used in [Hayes-LCN14], which
   calculated freq_est every T using the current E_N(E_T(OWD)).  Our
   tests show that this approximation of freq_est yields results that
   are almost identical to when the full calculation is performed
   every T.

3.2.5.  Packet Loss



   The proportion of packets lost over the period NT is used as a
   supplementary measure:

      pkt_loss = sum_NT(lost packets) / sum_NT(total packets)

   Note: When pkt_loss is low, it is very variable; however, when
   pkt_loss is high, it becomes a stable measure for making grouping
   decisions.

3.3.  Flow Grouping



3.3.1.  Flow-Grouping Algorithm



   The following grouping algorithm is RECOMMENDED for the use of SBD
   with coupled congestion control for RTP media [RTP-COUPLED-CC] and is
   sufficient and efficient for small to moderate numbers of flows.  For
   very large numbers of flows (e.g., hundreds), a more complex
   clustering algorithm may be substituted.

   Since no single metric is precise enough to group flows (due to
   noise), the algorithm uses multiple metrics.  Each metric offers a
   different "view" of the bottleneck link characteristics, and used
   together they enable a more precise grouping of flows than would
   otherwise be possible.




Hayes, et al.                 Experimental                     [Page 14]

RFC 8382                SBD for CCC for RTP Media              June 2018


   Flows determined to be transiting a bottleneck are successively
   divided into groups based on freq_est, var_est, skew_est, and
   pkt_loss.

   The first step is to determine which flows are transiting a
   bottleneck.  This is important, since if a flow is not transiting a
   bottleneck its delay-based metrics will not describe the bottleneck
   but will instead describe the "noise" from the rest of the path.
   Skewness, with the proportion of packet loss as a supplementary
   measure, is used to do this:

   1.  Grouping will be performed on flows that are inferred to be
       traversing a bottleneck by:

          skew_est < c_s

             || ( skew_est < c_h & PB ) || pkt_loss > p_l

       The parameter c_s controls how sensitive the mechanism is in
       detecting a bottleneck.  c_s = 0.0 was used in [Hayes-LCN14].  A
       value of c_s = 0.1 is a little more sensitive, and c_s = -0.1 is
       a little less sensitive.  c_h controls the hysteresis on flows
       that were grouped as transiting a bottleneck the previous time.
       If the test result is TRUE, PB=TRUE; otherwise, PB=FALSE.

   These flows (i.e., flows transiting a bottleneck) are then
   progressively divided into groups based on the freq_est, var_est, and
   skew_est summary statistics.  The process proceeds according to the
   following steps:

   2.  Group flows whose difference in sorted freq_est is less than a
       threshold:

          diff(freq_est) < p_f

   3.  Subdivide the groups obtained in step 2 by grouping flows whose
       difference in sorted E_M(var_est) (highest to lowest) is less
       than a threshold:

          diff(var_est) < (p_mad * var_est)

       The threshold, (p_mad * var_est), is with respect to the highest
       value in the difference.








Hayes, et al.                 Experimental                     [Page 15]

RFC 8382                SBD for CCC for RTP Media              June 2018


   4.  Subdivide the groups obtained in step 3 by grouping flows whose
       difference in sorted skew_est is less than a threshold:



          diff(skew_est) < p_s

   5.  When packet loss is high enough to be reliable (pkt_loss > p_l),
       subdivide the groups obtained in step 4 by grouping flows whose
       difference is less than a threshold:

          diff(pkt_loss) < (p_d * pkt_loss)

       The threshold, (p_d * pkt_loss), is with respect to the highest
       value in the difference.






































Hayes, et al.                 Experimental                     [Page 16]

RFC 8382                SBD for CCC for RTP Media              June 2018


   This procedure involves sorting estimates from highest to lowest.  It
   is simple to implement and is efficient for small numbers of flows
   (up to 10-20).  Figure 2 illustrates this algorithm.

                                        *********
                                        * Flows *
                                        ***.**.**
                                          /    '
                                         /     '--.
                                        /          \
                                   .---v--.    .----v---.
   1. Flows traversing             | Cong |    | UnCong |
      a bottleneck                 '-.--.-'    '--------'
                                    /    \
                                   /      \
                                  /        \
                              .--v--.       v-----.
   2. Divide by               | g_1 |  ...  | g_n |
      freq_est                '---.-.       '----..
                                 /   \          /  \
                                /     '--.     v    '------.
                               /          \                 \
                         .----v-.        .-v----.        .---v--.
   3. Divide by          | g_1a |  ...   | g_1z |   ...  | g_nz |
      var_est            '---.-.'        '-----..        '-.-.--'
                            /   \             /  \        /  |
                           /     '-----.     v    v      v   |
                          /             \                    |
                       .-v-----.       .-v-----.         .---v---.
   4. Divide by        | g_1ai |  ...  | g_1ax |   ...   | g_nzx |
      skew_est         '----.-.'       '------..         '-.-.---'
                           /   \             /  \         /  |
                          /     '--.        v    v       v   |
                         /          \                        |
                  .-----v--.       .-v------.           .----v---.
   5. Divide by   | g_1aiA |  ...  | g_1aiZ |    ...    | g_nzxZ |
      pkt_loss    '--------'       '--------'           '--------'
      (when applicable)

                         Simple grouping algorithm

                                 Figure 2









Hayes, et al.                 Experimental                     [Page 17]

RFC 8382                SBD for CCC for RTP Media              June 2018


3.3.2.  Using the Flow Group Signal



   Grouping decisions can be made every T from the second T; however,
   they will not attain their full design accuracy until after the
   2*Nth T interval.  We recommend that grouping decisions not be made
   until 2*M T intervals.

   Network conditions, and even the congestion controllers, can cause
   bottlenecks to fluctuate.  A coupled congestion controller MAY decide
   only to couple groups that remain stable, say grouped together 90% of
   the time, depending on its objectives.  Recommendations concerning
   this are beyond the scope of this document and will be specific to
   the coupled congestion controller's objectives.

4.  Enhancements to the Basic SBD Algorithm

   The SBD algorithm as specified in Section 3 was found to work well
   for a broad variety of conditions.  The following enhancements to the
   basic mechanisms have been found to significantly improve the
   algorithm's performance under some circumstances and SHOULD be
   implemented.  These "tweaks" are described separately to keep the
   main description succinct.

4.1.  Reducing Lag and Improving Responsiveness



   This section describes how to improve the responsiveness of the basic
   algorithm.

   Measurement-based shared bottleneck detection makes decisions in the
   present based on what has been measured in the past.  This means that
   there is always a lag in responding to changing conditions.  This
   mechanism is based on summary statistics taken over (N*T) seconds.
   This mechanism can be made more responsive to changing conditions by:

   1.  Reducing N and/or M, but at the expense of having metrics that
       are less accurate, and/or

   2.  Exploiting the fact that measurements that are more recent are
       more valuable than older measurements and weighting them
       accordingly.

   Although measurements that are more recent are more valuable, older
   measurements are still needed to gain an accurate estimate of the
   distribution descriptor we are measuring.  Unfortunately, the simple
   exponentially weighted moving average weights drop off too quickly
   for our requirements and have an infinite tail.  A simple linearly
   declining weighted moving average also does not provide enough weight
   to the measurements that are most recent.  We propose a piecewise



Hayes, et al.                 Experimental                     [Page 18]

RFC 8382                SBD for CCC for RTP Media              June 2018


   linear distribution of weights, such that the first section (samples
   1:F) is flat as in a simple moving average, and the second section
   (samples F+1:M) is linearly declining weights to the end of the
   averaging window.  We choose integer weights; this allows incremental
   calculation without introducing rounding errors.

4.1.1.  Improving the Response of the Skewness Estimate



   The weighted moving average for skew_est, based on skew_est as
   defined in Section 3.2.2, can be calculated as follows:

      skew_est = ((M-F+1)*sum(skew_base_T(1:F))

                      + sum([(M-F):1].*skew_base_T(F+1:M)))

                 / ((M-F+1)*sum(numsampT(1:F))

                      + sum([(M-F):1].*numsampT(F+1:M)))

   where numsampT is an array of the number of OWD samples in each T
   (i.e., num_T(OWD)), and numsampT(1) is the most recent;
   skew_base_T(1) is the most recent calculation of skew_base_T; 1:F
   refers to the integer values 1 through to F, and [(M-F):1] refers to
   an array of the integer values (M-F) declining through to 1; and ".*"
   is the array scalar dot product operator.

   To calculate this weighted skew_est incrementally:

   Notation:    F_ = flat portion, D_ = declining portion,
                W_ = weighted component

   Initialize:  sum_skewbase = 0, F_skewbase = 0, W_D_skewbase = 0

                skewbase_hist = buffer of length M, initialized to 0

                numsampT = buffer of length M, initialized to 0

   Steps per iteration:

   1.   old_skewbase = skewbase_hist(M)

   2.   old_numsampT = numsampT(M)

   3.   cycle(skewbase_hist)

   4.   cycle(numsampT)

   5.   numsampT(1) = num_T(OWD)





Hayes, et al.                 Experimental                     [Page 19]

RFC 8382                SBD for CCC for RTP Media              June 2018


   6.   skewbase_hist(1) = skew_base_T



   7.   F_skewbase = F_skewbase + skew_base_T - skewbase_hist(F+1)



   8.   W_D_skewbase = W_D_skewbase + (M-F)*skewbase_hist(F+1)
          - sum_skewbase



   9.   W_D_numsamp = W_D_numsamp + (M-F)*numsampT(F+1) - sum_numsamp
          + F_numsamp



   10.  F_numsamp = F_numsamp + numsampT(1) - numsampT(F+1)



   11.  sum_skewbase = sum_skewbase + skewbase_hist(F+1) - old_skewbase



   12.  sum_numsamp = sum_numsamp + numsampT(1) - old_numsampT



   13.  skew_est = ((M-F+1)*F_skewbase + W_D_skewbase) /
          ((M-F+1)*F_numsamp+W_D_numsamp)



   where cycle(...) refers to the operation on a cyclic buffer where the
   start of the buffer is now the next element in the buffer.

4.1.2.  Improving the Response of the Variability Estimate



   Similarly, the weighted moving average for var_est can be calculated
   as follows:

      var_est = ((M-F+1)*sum(var_base_T(1:F))

                     + sum([(M-F):1].*var_base_T(F+1:M)))

                / ((M-F+1)*sum(numsampT(1:F))

                     + sum([(M-F):1].*numsampT(F+1:M)))

   where numsampT is an array of the number of OWD samples in each T
   (i.e., num_T(OWD)), and numsampT(1) is the most recent;
   skew_base_T(1) is the most recent calculation of skew_base_T; 1:F
   refers to the integer values 1 through to F, and [(M-F):1] refers to
   an array of the integer values (M-F) declining through to 1; and ".*"
   is the array scalar dot product operator.  When removing oscillation
   noise (see Section 4.2), this calculation must be adjusted to allow
   for invalid var_base_T records.

   var_est can be calculated incrementally in the same way as skew_est
   as shown in Section 4.1.1.  However, note that the buffer numsampT is
   used for both calculations, so the operations on it should not be
   repeated.



Hayes, et al.                 Experimental                     [Page 20]

RFC 8382                SBD for CCC for RTP Media              June 2018


4.2.  Removing Oscillation Noise



   When a path has no bottleneck, var_est will be very small and the
   recorded significant mean crossings will be the result of path noise.
   Thus, up to N-1 meaningless mean crossings can be a source of error
   at the point where a link becomes a bottleneck and flows traversing
   it begin to be grouped.

   To remove this source of noise from freq_est:

   1.  Set the current var_base_T = NaN (a value representing an invalid
       record, i.e., Not a Number) for flows that are deemed to not be
       transiting a bottleneck by the first grouping test that is based
       on skew_est (see Section 3.3.1).

   2.  Then, var_est = sum_MT(var_base_T != NaN) / num_MT(OWD).

   3.  For freq_est, only record a significant mean crossing if a given
       flow is deemed to be transiting a bottleneck.

   These three changes can help to remove the non-bottleneck noise from
   freq_est.

5.  Measuring OWD

   This section discusses the OWD measurements required for this
   algorithm to detect shared bottlenecks.

   The SBD mechanism described in this document relies on differences
   between OWD measurements to avoid the practical problems with
   measuring absolute OWD (see [Hayes-LCN14], Section III.C).  Since all
   summary statistics are relative to the mean OWD and sender/receiver
   clock offsets should be approximately constant over the measurement
   periods, the offset is subtracted out in the calculation.

5.1.  Timestamp Resolution



   The SBD mechanism requires timing information precise enough to be
   able to make comparisons.  As a rule of thumb, the time resolution
   should be less than one hundredth of a typical path's range of
   delays.  In general, the coarser the time resolution, the more care
   that needs to be taken to ensure that rounding errors do not bias the
   skewness calculation.  Frequent timing information in millisecond
   resolution as described by [RTCP-CC-FEEDBACK] should be sufficient
   for the sender to calculate relative OWD.






Hayes, et al.                 Experimental                     [Page 21]

RFC 8382                SBD for CCC for RTP Media              June 2018


5.2.  Clock Skew



   Generally, sender and receiver clock skew will be too small to cause
   significant errors in the estimators.  skew_est and freq_est are the
   most sensitive to this type of noise due to their use of a mean OWD
   calculated over a longer interval.  In circumstances where clock skew
   is high, basing skew_est only on the previous T's mean and ignoring
   freq_est provide a noisier but reliable signal.

   A more sophisticated method is to estimate the effect the clock skew
   is having on the summary statistics and then adjust statistics
   accordingly.  There are a number of techniques in the literature,
   including [Zhang-Infocom02].

6.  Expected Feedback from Experiments

   The algorithm described in this memo has so far been evaluated using
   simulations and small-scale experiments.  Real network tests using
   RTP Media Congestion Avoidance Techniques (RMCAT) congestion control
   algorithms will help confirm the default parameter choice.  For
   example, the time interval T may need to be made longer if the packet
   rate is very low.  Implementers and testers are invited to document
   their findings in an Internet-Draft.

7.  IANA Considerations

   This document has no IANA actions.

8.  Security Considerations

   The security considerations of RFC 3550 [RFC3550], RFC 4585
   [RFC4585], and RFC 5124 [RFC5124] are expected to apply.

   Non-authenticated RTCP packets carrying OWD measurements, shared
   bottleneck indications, and/or summary statistics could allow
   attackers to alter the bottleneck-sharing characteristics for private
   gain or disruption of other parties' communication.  When using SBD
   for coupled congestion control as described in [RTP-COUPLED-CC], the
   security considerations of [RTP-COUPLED-CC] apply.












Hayes, et al.                 Experimental                     [Page 22]

RFC 8382                SBD for CCC for RTP Media              June 2018


9.  References

9.1.  Normative References



   [RFC2119]  Bradner, S., "Key words for use in RFCs to Indicate
              Requirement Levels", BCP 14, RFC 2119,
              DOI 10.17487/RFC2119, March 1997,
              <https://www.rfc-editor.org/info/rfc2119>.

   [RFC8174]  Leiba, B., "Ambiguity of Uppercase vs Lowercase in
              RFC 2119 Key Words", BCP 14, RFC 8174,
              DOI 10.17487/RFC8174, May 2017,
              <https://www.rfc-editor.org/info/rfc8174>.

9.2.  Informative References



   [Hayes-LCN14]
              Hayes, D., Ferlin, S., and M. Welzl, "Practical Passive
              Shared Bottleneck Detection using Shape Summary
              Statistics", Proc. IEEE Local Computer Networks (LCN),
              pp. 150-158, DOI 10.1109/LCN.2014.6925767, September 2014,
              <http://heim.ifi.uio.no/davihay/
              hayes14__pract_passiv_shared_bottl_detec-abstract.html>.

   [RFC3550]  Schulzrinne, H., Casner, S., Frederick, R., and V.
              Jacobson, "RTP: A Transport Protocol for Real-Time
              Applications", STD 64, RFC 3550, DOI 10.17487/RFC3550,
              July 2003, <https://www.rfc-editor.org/info/rfc3550>.

   [RFC4585]  Ott, J., Wenger, S., Sato, N., Burmeister, C., and J. Rey,
              "Extended RTP Profile for Real-time Transport Control
              Protocol (RTCP)-Based Feedback (RTP/AVPF)", RFC 4585,
              DOI 10.17487/RFC4585, July 2006,
              <https://www.rfc-editor.org/info/rfc4585>.

   [RFC5124]  Ott, J. and E. Carrara, "Extended Secure RTP Profile for
              Real-time Transport Control Protocol (RTCP)-Based Feedback
              (RTP/SAVPF)", RFC 5124, DOI 10.17487/RFC5124,
              February 2008, <https://www.rfc-editor.org/info/rfc5124>.

   [RFC6817]  Shalunov, S., Hazel, G., Iyengar, J., and M. Kuehlewind,
              "Low Extra Delay Background Transport (LEDBAT)", RFC 6817,
              DOI 10.17487/RFC6817, December 2012,
              <https://www.rfc-editor.org/info/rfc6817>.







Hayes, et al.                 Experimental                     [Page 23]

RFC 8382                SBD for CCC for RTP Media              June 2018


   [RFC7679]  Almes, G., Kalidindi, S., Zekauskas, M., and A. Morton,
              Ed., "A One-Way Delay Metric for IP Performance Metrics
              (IPPM)", STD 81, RFC 7679, DOI 10.17487/RFC7679,
              January 2016, <https://www.rfc-editor.org/info/rfc7679>.

   [RFC7680]  Almes, G., Kalidindi, S., Zekauskas, M., and A. Morton,
              Ed., "A One-Way Loss Metric for IP Performance Metrics
              (IPPM)", STD 82, RFC 7680, DOI 10.17487/RFC7680,
              January 2016, <https://www.rfc-editor.org/info/rfc7680>.

   [RTCP-CC-FEEDBACK]
              Sarker, Z., Perkins, C., Singh, V., and M. Ramalho,
              "RTP Control Protocol (RTCP) Feedback for Congestion
              Control", Work in Progress, draft-ietf-avtcore-cc-
              feedback-message-01, March 2018.

   [RTP-COUPLED-CC]
              Islam, S., Welzl, M., and S. Gjessing, "Coupled congestion
              control for RTP media", Work in Progress, draft-ietf-
              rmcat-coupled-cc-07, September 2017.

   [Zhang-Infocom02]
              Zhang, L., Liu, Z., and H. Xia, "Clock synchronization
              algorithms for network measurements", Proc. IEEE
              International Conference on Computer Communications
              (INFOCOM), pp. 160-169, DOI 10.1109/INFCOM.2002.1019257,
              September 2002.
























Hayes, et al.                 Experimental                     [Page 24]

RFC 8382                SBD for CCC for RTP Media              June 2018


Acknowledgments

   This work was partially funded by the European Community under its
   Seventh Framework Programme through the Reducing Internet Transport
   Latency (RITE) project (ICT-317700).  The views expressed are solely
   those of the authors.

Authors' Addresses



   David Hayes (editor)
   Simula Research Laboratory
   P.O. Box 134
   Lysaker  1325
   Norway

   Email: davidh@simula.no


   Simone Ferlin
   Simula Research Laboratory
   P.O. Box 134
   Lysaker  1325
   Norway

   Email: simone@ferlin.io


   Michael Welzl
   University of Oslo
   P.O. Box 1080 Blindern
   Oslo  N-0316
   Norway

   Email: michawe@ifi.uio.no


   Kristian Hiorth
   University of Oslo
   P.O. Box 1080 Blindern
   Oslo  N-0316
   Norway

   Email: kristahi@ifi.uio.no








Hayes, et al.                 Experimental                     [Page 25]