idnits 2.17.1 draft-ietf-rmcat-sbd-07.txt: Checking boilerplate required by RFC 5378 and the IETF Trust (see https://trustee.ietf.org/license-info): ---------------------------------------------------------------------------- No issues found here. Checking nits according to https://www.ietf.org/id-info/1id-guidelines.txt: ---------------------------------------------------------------------------- No issues found here. Checking nits according to https://www.ietf.org/id-info/checklist : ---------------------------------------------------------------------------- No issues found here. Miscellaneous warnings: ---------------------------------------------------------------------------- == The copyright year in the IETF Trust and authors Copyright Line does not match the current year -- The document date (June 8, 2017) is 2485 days in the past. Is this intentional? Checking references for intended status: Experimental ---------------------------------------------------------------------------- == Outdated reference: A later version (-04) exists of draft-dt-rmcat-feedback-message-02 == Outdated reference: A later version (-09) exists of draft-ietf-rmcat-coupled-cc-06 Summary: 0 errors (**), 0 flaws (~~), 3 warnings (==), 1 comment (--). Run idnits with the --verbose option for more detailed information about the items above. -------------------------------------------------------------------------------- 2 RTP Media Congestion Avoidance Techniques D. Hayes, Ed. 3 Internet-Draft S. Ferlin 4 Intended status: Experimental Simula Research Laboratory 5 Expires: December 10, 2017 M. Welzl 6 K. Hiorth 7 University of Oslo 8 June 8, 2017 10 Shared Bottleneck Detection for Coupled Congestion Control for RTP 11 Media. 12 draft-ietf-rmcat-sbd-07 14 Abstract 16 This document describes a mechanism to detect whether end-to-end data 17 flows share a common bottleneck. It relies on summary statistics 18 that are calculated based on continuous measurements and used as 19 input to a grouping algorithm that runs wherever the knowledge is 20 needed. This mechanism complements the coupled congestion control 21 mechanism in draft-ietf-rmcat-coupled-cc. 23 Status of This Memo 25 This Internet-Draft is submitted in full conformance with the 26 provisions of BCP 78 and BCP 79. 28 Internet-Drafts are working documents of the Internet Engineering 29 Task Force (IETF). Note that other groups may also distribute 30 working documents as Internet-Drafts. The list of current Internet- 31 Drafts is at http://datatracker.ietf.org/drafts/current/. 33 Internet-Drafts are draft documents valid for a maximum of six months 34 and may be updated, replaced, or obsoleted by other documents at any 35 time. It is inappropriate to use Internet-Drafts as reference 36 material or to cite them other than as "work in progress." 38 This Internet-Draft will expire on December 10, 2017. 40 Copyright Notice 42 Copyright (c) 2017 IETF Trust and the persons identified as the 43 document authors. All rights reserved. 45 This document is subject to BCP 78 and the IETF Trust's Legal 46 Provisions Relating to IETF Documents 47 (http://trustee.ietf.org/license-info) in effect on the date of 48 publication of this document. Please review these documents 49 carefully, as they describe your rights and restrictions with respect 50 to this document. Code Components extracted from this document must 51 include Simplified BSD License text as described in Section 4.e of 52 the Trust Legal Provisions and are provided without warranty as 53 described in the Simplified BSD License. 55 Table of Contents 57 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 3 58 1.1. The basic mechanism . . . . . . . . . . . . . . . . . . . 3 59 1.2. The signals . . . . . . . . . . . . . . . . . . . . . . . 3 60 1.2.1. Packet loss . . . . . . . . . . . . . . . . . . . . . 3 61 1.2.2. Packet delay . . . . . . . . . . . . . . . . . . . . 3 62 1.2.3. Path lag . . . . . . . . . . . . . . . . . . . . . . 4 63 2. Definitions . . . . . . . . . . . . . . . . . . . . . . . . . 4 64 2.1. Parameters and their effect . . . . . . . . . . . . . . . 7 65 2.2. Recommended parameter values . . . . . . . . . . . . . . 8 66 3. Mechanism . . . . . . . . . . . . . . . . . . . . . . . . . . 8 67 3.1. SBD feedback requirements . . . . . . . . . . . . . . . . 9 68 3.1.1. Feedback when all the logic is placed at the sender . 9 69 3.1.2. Feedback when the statistics are calculated at the 70 receiver and SBD performed at the sender . . . . . . 10 71 3.1.3. Feedback when bottlenecks can be determined at both 72 senders and receivers . . . . . . . . . . . . . . . . 11 73 3.2. Key metrics and their calculation . . . . . . . . . . . . 11 74 3.2.1. Mean delay . . . . . . . . . . . . . . . . . . . . . 11 75 3.2.2. Skewness estimate . . . . . . . . . . . . . . . . . . 11 76 3.2.3. Variability estimate . . . . . . . . . . . . . . . . 12 77 3.2.4. Oscillation estimate . . . . . . . . . . . . . . . . 12 78 3.2.5. Packet loss . . . . . . . . . . . . . . . . . . . . . 13 79 3.3. Flow Grouping . . . . . . . . . . . . . . . . . . . . . . 13 80 3.3.1. Flow grouping algorithm . . . . . . . . . . . . . . . 13 81 3.3.2. Using the flow group signal . . . . . . . . . . . . . 16 82 4. Enhancements to the basic SBD algorithm . . . . . . . . . . . 17 83 4.1. Reducing lag and improving responsiveness . . . . . . . . 17 84 4.1.1. Improving the response of the skewness estimate . . . 18 85 4.1.2. Improving the response of the variability estimate . 20 86 4.2. Removing oscillation noise . . . . . . . . . . . . . . . 20 87 5. Measuring OWD . . . . . . . . . . . . . . . . . . . . . . . . 21 88 5.1. Time stamp resolution . . . . . . . . . . . . . . . . . . 21 89 5.2. Clock skew . . . . . . . . . . . . . . . . . . . . . . . 21 90 6. Expected feedback from experiments . . . . . . . . . . . . . 21 91 7. Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . 22 92 8. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 22 93 9. Security Considerations . . . . . . . . . . . . . . . . . . . 22 94 10. Change history . . . . . . . . . . . . . . . . . . . . . . . 22 95 11. References . . . . . . . . . . . . . . . . . . . . . . . . . 23 96 11.1. Normative References . . . . . . . . . . . . . . . . . . 23 97 11.2. Informative References . . . . . . . . . . . . . . . . . 23 98 Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . 25 100 1. Introduction 102 In the Internet, it is not normally known if flows (e.g., TCP 103 connections or UDP data streams) traverse the same bottlenecks. Even 104 flows that have the same sender and receiver may take different paths 105 and may or may not share a bottleneck. Flows that share a bottleneck 106 link usually compete with one another for their share of the 107 capacity. This competition has the potential to increase packet loss 108 and delays. This is especially relevant for interactive applications 109 that communicate simultaneously with multiple peers (such as multi- 110 party video). For RTP media applications such as RTCWEB, 111 [I-D.ietf-rmcat-coupled-cc] describes a scheme that combines the 112 congestion controllers of flows in order to honor their priorities 113 and avoid unnecessary packet loss as well as delay. This mechanism 114 relies on some form of Shared Bottleneck Detection (SBD); here, a 115 measurement-based SBD approach is described. 117 1.1. The basic mechanism 119 The mechanism groups flows that have similar statistical 120 characteristics together. Section 3.3.1 describes a simple method 121 for achieving this, however, a major part of this draft is concerned 122 with collecting suitable statistics for this purpose. 124 1.2. The signals 126 The current Internet is unable to explicitly inform endpoints as to 127 which flows share bottlenecks, so endpoints need to infer this from 128 whatever information is available to them. The mechanism described 129 here currently utilizes packet loss and packet delay, but is not 130 restricted to these. As ECN becomes more prevalent it too will 131 become a valuable base signal. 133 1.2.1. Packet loss 135 Packet loss is often a relatively rare signal. Therefore, on its own 136 it is of limited use for SBD, however, it is a valuable supplementary 137 measure when it is more prevalent. 139 1.2.2. Packet delay 141 End-to-end delay measurements include noise from every device along 142 the path in addition to the delay perturbation at the bottleneck 143 device. The noise is often significantly increased if the round-trip 144 time is used. The cleanest signal is obtained by using One-Way-Delay 145 (OWD). 147 Measuring absolute OWD is difficult since it requires both the sender 148 and receiver clocks to be synchronized. However, since the 149 statistics being collected are relative to the mean OWD, a relative 150 OWD measurement is sufficient. Clock skew is not usually significant 151 over the time intervals used by this SBD mechanism (see [RFC6817] A.2 152 for a discussion on clock skew and OWD measurements). However, in 153 circumstances where it is significant, Section 5.2 outlines a way of 154 adjusting the calculations to cater for it. 156 Each packet arriving at the bottleneck buffer may experience very 157 different queue lengths, and therefore different waiting times. A 158 single OWD sample does not, therefore, characterize the path well. 159 However, multiple OWD measurements do reflect the distribution of 160 delays experienced at the bottleneck. 162 1.2.3. Path lag 164 Flows that share a common bottleneck may traverse different paths, 165 and these paths will often have different base delays. This makes it 166 difficult to correlate changes in delay or loss. This technique uses 167 the long term shape of the delay distribution as a base for 168 comparison to counter this. 170 2. Definitions 172 The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", 173 "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this 174 document are to be interpreted as described in RFC 2119 [RFC2119]. 176 Acronyms used in this document: 178 OWD -- One Way Delay 180 MAD -- Mean Absolute Deviation 182 RTT -- Round Trip Time 184 SBD -- Shared Bottleneck Detection 186 Conventions used in this document: 188 T -- the base time interval over which measurements are 189 made. 191 N -- the number of base time, T, intervals used in some 192 calculations. 194 M -- the number of base time, T, intervals used in some 195 calculations. 197 sum_T(...) -- summation of all the measurements of the variable 198 in parentheses taken over the interval T 200 sum(...) -- summation of terms of the variable in parentheses 202 sum_N(...) -- summation of N terms of the variable in parentheses 204 sum_NT(...) -- summation of all measurements taken over the 205 interval N*T 207 sum_MT(...) -- summation of all measurements taken over the 208 interval M*T 210 E_T(...) -- the expectation or mean of the measurements of the 211 variable in parentheses over T 213 E_N(...) -- the expectation or mean of the last N values of the 214 variable in parentheses 216 E_M(...) -- the expectation or mean of the last M values of the 217 variable in parentheses, where M <= N. 219 max_T(...) -- the maximum recorded measurement of the variable in 220 parentheses taken over the interval T 222 min_T(...) -- the minimum recorded measurement of the variable in 223 parentheses taken over the interval T 225 num_T(...) -- the count of measurements of the variable in 226 parentheses taken in the interval T 228 num_VM(...) -- the count of valid values of the variable in 229 parentheses given M records 231 PB -- a boolean variable indicating the particular flow 232 was identified transiting a bottleneck in the 233 previous interval T (i.e. Previously Bottleneck) 235 skew_est -- a measure of skewness in a OWD distribution. 237 skew_base_T -- a variable used as an intermediate step in 238 calculating skew_est. 240 var_est -- a measure of variability in OWD measurements. 242 var_base_T -- a variable used as an intermediate step in 243 calculating var_est. 245 freq_est -- a measure of low frequency oscillation in the OWD 246 measurements. 248 p_l, p_f, p_mad, c_s, c_h, p_s, p_d, p_v -- various thresholds 249 used in the mechanism 251 M and F -- number of values related to N 253 . 255 2.1. Parameters and their effect 257 T T should be long enough so that there are enough packets 258 received during T for a useful estimate of short term mean 259 OWD and variation statistics. Making T too large can limit 260 the efficacy of freq_est. It will also increase the response 261 time of the mechanism. Making T too small will make the 262 metrics noisier. 264 N & M N should be large enough to provide a stable estimate of 265 oscillations in OWD. Usually M=N, though having M mean_delay) skew_base_T-- 496 The mean_delay does not include the mean of the current T interval to 497 enable it to be calculated iteratively. 499 skew_est = sum_MT(skew_base_T)/num_MT(OWD) 501 where skew_est is a number between -1 and 1 503 Note: Care must be taken when implementing the comparisons to ensure 504 that rounding does not bias skew_est. It is important that the mean 505 is calculated with a higher precision than the samples. 507 3.2.3. Variability estimate 509 Mean Absolute Deviation (MAD) delay is a robust variability measure 510 that copes well with different send rates. It can be implemented in 511 an online manner as follows: 513 var_base_T = sum_T(|OWD - E_T(OWD)|) 515 where 517 |x| is the absolute value of x 519 E_T(OWD) is the mean OWD calculated in the previous T 521 var_est = MAD_MT = sum_MT(var_base_T)/num_MT(OWD) 523 For calculation of freq_est p_v=0.7 525 For the grouping threshold p_mad=0.1 527 3.2.4. Oscillation estimate 529 An estimate of the low frequency oscillation of the delay signal is 530 calculated by counting and normalizing the significant mean, 531 E_T(OWD), crossings of mean_delay: 533 freq_est = number_of_crossings / N 534 where we define a significant mean crossing as a crossing that 535 extends p_v * var_est from mean_delay. In our experiments we 536 have found that p_v = 0.7 is a good value. 538 Freq_est is a number between 0 and 1. Freq_est can be approximated 539 incrementally as follows: 541 With each new calculation of E_T(OWD) a decision is made as to 542 whether this value of E_T(OWD) significantly crosses the current 543 long term mean, mean_delay, with respect to the previous 544 significant mean crossing. 546 A cyclic buffer, last_N_crossings, records a 1 if there is a 547 significant mean crossing, otherwise a 0. 549 The counter, number_of_crossings, is incremented when there is a 550 significant mean crossing and decremented when a non-zero value is 551 removed from the last_N_crossings. 553 This approximation of freq_est was not used in [Hayes-LCN14], which 554 calculated freq_est every T using the current E_N(E_T(OWD)). Our 555 tests show that this approximation of freq_est yields results that 556 are almost identical to when the full calculation is performed every 557 T. 559 3.2.5. Packet loss 561 The proportion of packets lost over the period NT is used as a 562 supplementary measure: 564 pkt_loss = sum_NT(lost packets) / sum_NT(total packets) 566 Note: When pkt_loss is small it is very variable, however, when 567 pkt_loss is high it becomes a stable measure for making grouping 568 decisions. 570 3.3. Flow Grouping 572 3.3.1. Flow grouping algorithm 574 The following grouping algorithm is RECOMMENDED for SBD in the RMCAT 575 context and is sufficient and efficient for small to moderate numbers 576 of flows. For very large numbers of flows (e.g. hundreds), a more 577 complex clustering algorithm may be substituted. 579 Since no single metric is precise enough to group flows (due to 580 noise), the algorithm uses multiple metrics. Each metric offers a 581 different "view" of the bottleneck link characteristics, and used 582 together they enable a more precise grouping of flows than would 583 otherwise be possible. 585 Flows determined to be transiting a bottleneck are successively 586 divided into groups based on freq_est, var_est, skew_est and 587 pkt_loss. 589 The first step is to determine which flows are transiting a 590 bottleneck. This is important, since if a flow is not transiting a 591 bottleneck its delay based metrics will not describe the bottleneck, 592 but the "noise" from the rest of the path. Skewness, with proportion 593 of packet loss as a supplementary measure, is used to do this: 595 1. Grouping will be performed on flows that are inferred to be 596 traversing a bottleneck by: 598 skew_est < c_s 600 || ( skew_est < c_h & PB ) || pkt_loss > p_l 602 The parameter c_s controls how sensitive the mechanism is in 603 detecting a bottleneck. C_s = 0.0 was used in [Hayes-LCN14]. A 604 value of c_s = 0.05 is a little more sensitive, and c_s = -0.05 is a 605 little less sensitive. C_h controls the hysteresis on flows that 606 were grouped as transiting a bottleneck last time. If the test 607 result is TRUE, PB=TRUE, otherwise PB=FALSE. 609 These flows, flows transiting a bottleneck, are then progressively 610 divided into groups based on the freq_est, var_est, and skew_est 611 summary statistics. The process proceeds according to the following 612 steps: 614 2. Group flows whose difference in sorted freq_est is less than a 615 threshold: 617 diff(freq_est) < p_f 619 3. Subdivide the groups obtained in 2. by grouping flows whose 620 difference in sorted E_M(var_est) (highest to lowest) is less 621 than a threshold: 623 diff(var_est) < (p_mad * var_est) 625 The threshold, (p_mad * var_est), is with respect to the highest 626 value in the difference. 628 4. Subdivide the groups obtained in 3. by grouping flows whose 629 difference in sorted skew_est is less than a threshold: 631 diff(skew_est) < p_s 633 5. When packet loss is high enough to be reliable (pkt_loss > p_l), 634 Subdivide the groups obtained in 4. by grouping flows whose 635 difference is less than a threshold 637 diff(pkt_loss) < (p_d * pkt_loss) 639 The threshold, (p_d * pkt_loss), is with respect to the highest 640 value in the difference. 642 This procedure involves sorting estimates from highest to lowest. It 643 is simple to implement, and efficient for small numbers of flows (up 644 to 10-20).Figure 2 illustrates this algorithm 645 ********* 646 * Flows * 647 ***.**.** 648 / ' 649 / '--. 650 / \ 651 .---v--. .----v---. 652 1. Flows traversing | Cong | | UnCong | 653 a bottleneck '-.--.-' '--------' 654 / \ 655 / \ 656 / \ 657 .--v--. v-----. 658 2. Divide by | g_1 | ... | g_n | 659 freq_est '---.-. '----.. 660 / \ / \ 661 / '--. v '------. 662 / \ \ 663 .----v-. .-v----. .---v--. 664 3. Divide by | g_1a | ... | g_1z | ... | g_nz | 665 var_est '---.-.' '-----.. '-.-.--' 666 / \ / \ / | 667 / '-----. v v v | 668 / \ | 669 .-v-----. .-v-----. .---v---. 670 4. Divide by | g_1ai | ... | g_1ax | ... | g_nzx | 671 skew_est '----.-.' '------.. '-.-.---' 672 / \ / \ / | 673 / '--. v v v | 674 / \ | 675 .-----v--. .-v------. .----v---. 676 5. Divide by | g_1aiA | ... | g_1aiZ | ... | g_nzxZ | 677 pkt_loss '--------' '--------' '--------' 678 (when applicable) 680 Simple grouping algorithm. 682 Figure 2 684 3.3.2. Using the flow group signal 686 Grouping decisions can be made every T from the second T, however 687 they will not attain their full design accuracy until after the 688 2*N'th T interval. We recommend that grouping decisions are not made 689 until 2*M T intervals. 691 Network conditions, and even the congestion controllers, can cause 692 bottlenecks to fluctuate. A coupled congestion controller MAY decide 693 only to couple groups that remain stable, say grouped together 90% of 694 the time, depending on its objectives. Recommendations concerning 695 this are beyond the scope of this document and will be specific to 696 the coupled congestion controllers objectives. 698 4. Enhancements to the basic SBD algorithm 700 The SBD algorithm as specified in Section 3 was found to work well 701 for a broad variety of conditions. The following enhancements to the 702 basic mechanisms have been found to significantly improve the 703 algorithm's performance under some circumstances and SHOULD be 704 implemented. These "tweaks" are described separately to keep the 705 main description succinct. 707 4.1. Reducing lag and improving responsiveness 709 This section describes how to improve the responsiveness of the basic 710 algorithm. 712 Measurement based shared bottleneck detection makes decisions in the 713 present based on what has been measured in the past. This means that 714 there is always a lag in responding to changing conditions. This 715 mechanism is based on summary statistics taken over (N*T) seconds. 716 This mechanism can be made more responsive to changing conditions by: 718 1. Reducing N and/or M -- but at the expense of having less accurate 719 metrics, and/or 721 2. Exploiting the fact that more recent measurements are more 722 valuable than older measurements and weighting them accordingly. 724 Although more recent measurements are more valuable, older 725 measurements are still needed to gain an accurate estimate of the 726 distribution descriptor we are measuring. Unfortunately, the simple 727 exponentially weighted moving average weights drop off too quickly 728 for our requirements and have an infinite tail. A simple linearly 729 declining weighted moving average also does not provide enough weight 730 to the most recent measurements. We propose a piecewise linear 731 distribution of weights, such that the first section (samples 1:F) is 732 flat as in a simple moving average, and the second section (samples 733 F+1:M) is linearly declining weights to the end of the averaging 734 window. We choose integer weights, which allows incremental 735 calculation without introducing rounding errors. 737 4.1.1. Improving the response of the skewness estimate 739 The weighted moving average for skew_est, based on skew_est in 740 Section 3.2.2, can be calculated as follows: 742 skew_est = ((M-F+1)*sum(skew_base_T(1:F)) 744 + sum([(M-F):1].*skew_base_T(F+1:M))) 746 / ((M-F+1)*sum(numsampT(1:F)) 748 + sum([(M-F):1].*numsampT(F+1:M))) 750 where numsampT is an array of the number of OWD samples in each T 751 (i.e. num_T(OWD)), and numsampT(1) is the most recent; skew_base_T(1) 752 is the most recent calculation of skew_base_T; 1:F refers to the 753 integer values 1 through to F, and [(M-F):1] refers to an array of 754 the integer values (M-F) declining through to 1; and ".*" is the 755 array scalar dot product operator. 757 To calculate this weighted skew_est incrementally: 759 Notation: F_ - flat portion, D_ - declining portion, W_ - weighted 760 component 762 Initialize: sum_skewbase = 0, F_skewbase=0, W_D_skewbase=0 764 skewbase_hist = buffer length M initialize to 0 766 numsampT = buffer length M initialized to 0 768 Steps per iteration: 770 1. old_skewbase = skewbase_hist(M) 772 2. old_numsampT = numsampT(M) 774 3. cycle(skewbase_hist) 776 4. cycle(numsampT) 778 5. numsampT(1) = num_T(OWD) 780 6. skewbase_hist(1) = skew_base_T 782 7. F_skewbase = F_skewbase + skew_base_T - skewbase_hist(F+1) 784 8. W_D_skewbase = W_D_skewbase + (M-F)*skewbase_hist(F+1) 785 - sum_skewbase 787 9. W_D_numsamp = W_D_numsamp + (M-F)*numsampT(F+1) - sum_numsamp 788 + F_numsamp 790 10. F_numsamp = F_numsamp + numsampT(1) - numsampT(F+1) 792 11. sum_skewbase = sum_skewbase + skewbase_hist(F+1) - old_skewbase 794 12. sum_numsamp = sum_numsamp + numsampT(1) - old_numsampT 796 13. skew_est = ((M-F+1)*F_skewbase + W_D_skewbase) / 797 ((M-F+1)*F_numsamp+W_D_numsamp) 799 Where cycle(....) refers to the operation on a cyclic buffer where 800 the start of the buffer is now the next element in the buffer. 802 4.1.2. Improving the response of the variability estimate 804 Similarly the weighted moving average for var_est can be calculated 805 as follows: 807 var_est = ((M-F+1)*sum(var_base_T(1:F)) 809 + sum([(M-F):1].*var_base_T(F+1:M))) 811 / ((M-F+1)*sum(numsampT(1:F)) 813 + sum([(M-F):1].*numsampT(F+1:M))) 815 where numsampT is an array of the number of OWD samples in each T 816 (i.e. num_T(OWD)), and numsampT(1) is the most recent; skew_base_T(1) 817 is the most recent calculation of skew_base_T; 1:F refers to the 818 integer values 1 through to F, and [(M-F):1] refers to an array of 819 the integer values (M-F) declining through to 1; and ".*" is the 820 array scalar dot product operator. When removing oscillation noise 821 (see Section 4.2) this calculation must be adjusted to allow for 822 invalid var_base_T records. 824 Var_est can be calculated incrementally in the same way as skew_est 825 in Section 4.1.1. However, note that the buffer numsampT is used for 826 both calculations so the operations on it should not be repeated. 828 4.2. Removing oscillation noise 830 When a path has no bottleneck, var_est will be very small and the 831 recorded significant mean crossings will be the result of path noise. 832 Thus up to N-1 meaningless mean crossings can be a source of error at 833 the point a link becomes a bottleneck and flows traversing it begin 834 to be grouped. 836 To remove this source of noise from freq_est: 838 1. Set the current var_base_T = NaN (a value representing an invalid 839 record, i.e. Not a Number) for flows that are deemed to not be 840 transiting a bottleneck by the first skew_est based grouping test 841 (see Section 3.3.1). 843 2. Then var_est = sum_MT(var_base_T != NaN) / num_MT(OWD) 845 3. For freq_est, only record a significant mean crossing if flow 846 deemed to be transiting a bottleneck. 848 These three changes can help to remove the non-bottleneck noise from 849 freq_est. 851 5. Measuring OWD 853 This section discusses the OWD measurements required for this 854 algorithm to detect shared bottlenecks. 856 The SBD mechanism described in this document relies on differences 857 between OWD measurements to avoid the practical problems with 858 measuring absolute OWD (see [Hayes-LCN14] section IIIC). Since all 859 summary statistics are relative to the mean OWD and sender/receiver 860 clock offsets should be approximately constant over the measurement 861 periods, the offset is subtracted out in the calculation. 863 5.1. Time stamp resolution 865 The SBD mechanism requires timing information precise enough to be 866 able to make comparisons. As a rule of thumb, the time resolution 867 should be less than one hundredth of a typical path's range of 868 delays. In general, the coarser the time resolution, the more care 869 that needs to be taken to ensure rounding errors do not bias the 870 skewness calculation. Time stamp resolution such as that described 871 by [I-D.dt-rmcat-feedback-message] should be sufficient. 873 5.2. Clock skew 875 Generally sender and receiver clock skew will be too small to cause 876 significant errors in the estimators. Skew_est and freq_est are the 877 most sensitive to this type of noise due to their use of a mean OWD 878 calculated over a longer interval. In circumstances where clock skew 879 is high, basing skew_est only on the previous T's mean and ignoring 880 freq_est provides a noisier but reliable signal. 882 A more sophisticated method is to estimate the effect the clock skew 883 is having on the summary statistics, and then adjust statistics 884 accordingly. There are a number of techniques in the literature, 885 including [Zhang-Infocom02]. 887 6. Expected feedback from experiments 889 The algorithm described in this memo has so far been evaluated using 890 simulations. Real network tests using the proposed congestion 891 control algorithms will help confirm the default parameter choice. 892 For example, the time interval T may need to be made longer if the 893 packet rate is very low. Implementers and testers are invited to 894 document their findings in an Internet draft. 896 7. Acknowledgments 898 This work was part-funded by the European Community under its Seventh 899 Framework Programme through the Reducing Internet Transport Latency 900 (RITE) project (ICT-317700). The views expressed are solely those of 901 the authors. 903 8. IANA Considerations 905 This memo includes no request to IANA. 907 9. Security Considerations 909 The security considerations of RFC 3550 [RFC3550], RFC 4585 910 [RFC4585], and RFC 5124 [RFC5124] are expected to apply. 912 Non-authenticated RTCP packets carrying OWD measurements, shared 913 bottleneck indications, and/or summary statistics could allow 914 attackers to alter the bottleneck sharing characteristics for private 915 gain or disruption of other parties communication. 917 10. Change history 919 Changes made to this document: 921 WG-06->WG-07 : Updates addressing 922 https://mailarchive.ietf.org/arch/msg/ 923 rmcat/80B6q4nI7carGcf_ddBwx7nKvOw. Mainly 924 clarifications. Figure 2 to supplement grouping 925 algorithm description. 927 WG-05->WG-06 : Updates addressing WG reviews 928 https://mailarchive.ietf.org/arch/msg/rmcat/- 929 1JdrTMq1Y5T6ZNlOkrQJQ27TzE and 930 https://mailarchive.ietf.org/arch/msg/rmcat/ 931 eI2Q1f8NL2SxbJgjFLR4_rEmJ_g. This has mainly 932 involved minor clarifications, including the moving 933 of 3.4.1 and 3.5 into the new Section 4, and 3.4.1 934 into Section 5 936 WG-04->WG-05 : Fix ToC formatting. Add section on expected 937 feedback from experiments replacing short section 938 on implementation status. Added comment on ECN as 939 a signal. Clarification of lost packet signaling. 940 Change term "draft" to "document" where 941 appropriate. American spelling. Some tightening 942 of the text. 944 WG-03->WG-04 : Add M to terminology table, suggest skew_est based 945 on previous T and no freq_est in clock skew 946 section, feedback requirements as a separate sub 947 section. 949 WG-02->WG-03 : Correct misspelled author 951 WG-01->WG-02 : Removed ambiguity associated with the term 952 "congestion". Expanded the description of 953 initialization messages. Removed PDV metric. 954 Added description of incremental weighted metric 955 calculations for skew_est. Various clarifications 956 based on implementation work. Fixed typos and 957 tuned parameters. 959 WG-00->WG-01 : Moved unbiased skew section to replace skew 960 estimate, more robust variability estimator, the 961 term variance replaced with variability, clock 962 drift term corrected to clock skew, revision to 963 clock skew section with a place holder, description 964 of parameters. 966 02->WG-00 : Fixed missing 0.5 in 3.3.2 and missing brace in 967 3.3.3 969 01->02 : New section describing improvements to the key 970 metric calculations that help to remove noise, 971 bias, and reduce lag. Some revisions to the 972 notation to make it clearer. Some tightening of 973 the thresholds. 975 00->01 : Revisions to terminology for clarity 977 11. References 979 11.1. Normative References 981 [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate 982 Requirement Levels", BCP 14, RFC 2119, 983 DOI 10.17487/RFC2119, March 1997, 984 . 986 11.2. Informative References 988 [Hayes-LCN14] 989 Hayes, D., Ferlin, S., and M. Welzl, "Practical Passive 990 Shared Bottleneck Detection using Shape Summary 991 Statistics", Proc. the IEEE Local Computer Networks 992 (LCN) pp150-158, September 2014, 993 . 996 [I-D.dt-rmcat-feedback-message] 997 Sarker, Z., Perkins, C., Singh, V., and M. Ramalho, "RTP 998 Control Protocol (RTCP) Feedback for Congestion Control", 999 draft-dt-rmcat-feedback-message-02 (work in progress), May 1000 2017. 1002 [I-D.ietf-rmcat-coupled-cc] 1003 Islam, S., Welzl, M., and S. Gjessing, "Coupled congestion 1004 control for RTP media", draft-ietf-rmcat-coupled-cc-06 1005 (work in progress), March 2017. 1007 [RFC3550] Schulzrinne, H., Casner, S., Frederick, R., and V. 1008 Jacobson, "RTP: A Transport Protocol for Real-Time 1009 Applications", STD 64, RFC 3550, DOI 10.17487/RFC3550, 1010 July 2003, . 1012 [RFC4585] Ott, J., Wenger, S., Sato, N., Burmeister, C., and J. Rey, 1013 "Extended RTP Profile for Real-time Transport Control 1014 Protocol (RTCP)-Based Feedback (RTP/AVPF)", RFC 4585, 1015 DOI 10.17487/RFC4585, July 2006, 1016 . 1018 [RFC5124] Ott, J. and E. Carrara, "Extended Secure RTP Profile for 1019 Real-time Transport Control Protocol (RTCP)-Based Feedback 1020 (RTP/SAVPF)", RFC 5124, DOI 10.17487/RFC5124, February 1021 2008, . 1023 [RFC6817] Shalunov, S., Hazel, G., Iyengar, J., and M. Kuehlewind, 1024 "Low Extra Delay Background Transport (LEDBAT)", RFC 6817, 1025 DOI 10.17487/RFC6817, December 2012, 1026 . 1028 [Zhang-Infocom02] 1029 Zhang, L., Liu, Z., and H. Xia, "Clock synchronization 1030 algorithms for network measurements", Proc. the IEEE 1031 International Conference on Computer Communications 1032 (INFOCOM) pp160-169, September 2002, 1033 . 1035 Authors' Addresses 1037 David Hayes (editor) 1038 Simula Research Laboratory 1039 P.O. Box 134 1040 Lysaker 1325 1041 Norway 1043 Phone: +47 2284 5566 1044 Email: davidh@simula.no 1046 Simone Ferlin 1047 Simula Research Laboratory 1048 P.O.Box 134 1049 Lysaker 1325 1050 Norway 1052 Phone: +47 4072 0702 1053 Email: ferlin@simula.no 1055 Michael Welzl 1056 University of Oslo 1057 PO Box 1080 Blindern 1058 Oslo N-0316 1059 Norway 1061 Phone: +47 2285 2420 1062 Email: michawe@ifi.uio.no 1064 Kristian Hiorth 1065 University of Oslo 1066 PO Box 1080 Blindern 1067 Oslo N-0316 1068 Norway 1070 Email: kristahi@ifi.uio.no