ATIS 0800005-2006 IPTV Packet Loss Issue Report.pdf

1、 ATIS-0800005 IPTV PACKET LOSS ISSUE REPORT The Alliance for Telecommunication Industry Solutions (ATIS) is a technical planning and standards development organization that is committed to rapidly developing and promoting technical and operations standards for the communications and related informat

2、ion technologies industry worldwide using a pragmatic, flexible and open approach. Over 1,100 participants from more than 350 communications companies are active in ATIS 23 industry committees and its Incubator Solutions Program. NOTE - The users attention is called to the possibility that complianc

3、e with this standard may require use of an invention covered by patent rights. By publication of this standard, no position is taken with respect to whether use of an invention covered by patent rights will be required, and if any such use is required no position is taken regarding the validity of t

4、his claim or any patent rights in connection therewith. ATIS-0800005, IPTV Packet Loss Issue Report Is an ATIS standard developed by the Architecture (ARCH) Task Force under the ATIS IPTV Interoperability Forum (IIF). Published by Alliance for Telecommunications Industry Solutions 1200 G Street, NW,

5、 Suite 500 Washington, DC 20005 Copyright 2007 by Alliance for Telecommunications Industry Solutions All rights reserved. No part of this publication may be reproduced in any form, in an electronic retrieval system or otherwise, without the prior written permission of the publisher. For information

6、contact ATIS at 202.628.6380. ATIS is online at . Printed in the United States of America. ATIS-0800005 ATIS Standard on IPTV Packet Loss Issue Report Secretariat Alliance for Telecommunications Industry Solutions Approved December 2006 Abstract The IPTV Packet Loss Issue report is a technical repor

7、t that explores a range of potential solutions and makes recommendations regarding their applicability for an IPTV service. CONTRIBUTORS: Dipan Patel, Accenture TF Co-Chair Uche Ojeh, Accenture Kevin Schneider, ADTRAN Donald Crowe, Alcatel-Lucent Rajesh Jagannathan, Alcatel-Lucent Randy Sharpe, Alca

8、tel-Lucent TF Co-Chair Rick Townsend, Alcatel-Lucent Tim Barrett, Alcatel-Lucent Amit Kleinmann, Amdocs Al Morton, AT Electrical interference in the home network (e.g., Wireless); Packet discard mechanisms in network equipment; and Consumer Electronics (CE) network equipment response to transient co

9、mmercial power events. Common layer 2 protocols, such as Ethernet, enforce a packet discard if there is a discrepancy in the Frame Check Sequence. This behavior essentially magnifies a single bit error into a packet loss event. Other protocols (such as ATM) provide a mechanism to forward the errored

10、 packets allowing the terminal equipment to attempt some error recovery. Packet discard may also occur due to transient buffer overloads in network equipment. The buffer occupancy is to some extent controllable by the operator given adequate information about the traffic characteristics. There may b

11、e disadvantages to attempting to manage potential for buffer overload events, such as capital inefficiency, from running at low link and node utilizations. Furthermore, even in networks with low average utilization there can be wide variations in the instantaneous traffic load over both short and lo

12、ng periods of time. Thus it can be expected that such transient congestion events can be observed even in well-engineered networks. Such congestion losses should not be confused with persistent congestion that will occur if a link is permanently oversubscribed. It shall be assumed that a service pro

13、vider needs to adequately engineer for the services provided on their network, so that such persistent congestions should not occur and thus not require further consideration in this document. Video can be encoded in an open loop with a fixed quantization scale resulting in Variable Bit Rate (VBR) t

14、raffic with relatively constant quality, or encoded at a relatively constant bit rate with variable quality Lakshman. Current commercial digitized video packet formats have been optimized for dedicated transmission channels (e.g., satellite wavelengths) by various adaptation mechanisms to shape the

15、traffic to a relatively constant rate. Buffer overloads are not expected to be a significant issue with constant rate streams. In order to adequately dimension buffers for the statistical multiplexing of VBR video streams, the traffic statistics of the VBR streams need to be well- characterized. Whi

16、le statistical multiplexing gains may not be significant on a per subscriber interface (with . ATIS 800003 ATIS-0800003, ATIS IPTV Architecture Roadmap, August 2006. Mosko2000 M. Mosko, J.J. Garcia-Luna-Aceves. Analysis of Packet Loss Correlation in FEC-Enhanced Multicast Trees, Proceedings of IEEE

17、ICNP, 2000. 20 Annex A Models for DSL Link Packet Loss A.1 INTRODUCTION Once a link is set up and communicating successfully, packets are lost primarily either because of network congestion, or because of transmission errors. Transmission loss models are presented for uncorrectable errors from impul

18、se noise on an ADSL/2/2+ or VDSL2 line. This work was previously presented to the DSL Forum 1. A. 2 TRANSMISSION ERRORS: ADSL2+ AND VDSL2 Copper access links are known to occasionally experience uncorrectable burst errors from impulse noise. Impulse noise is short duration (tens of microseconds) bur

19、sts of high-power noise, often caused by transients on nearby powerlines. For Internet access, TCP/IP retransmits correct impulse noise errors, but video service may not have such capability. Random errors introduced by stationary noise are usually correctable on metallic lines, and optical lines us

20、ually have a very low error rate, so these are ignored and only impulse noise is considered here. Impulse noise is mitigated in ADSL, ADSL2, ADSL2+, VDSL, and VDSL2 by using interleaved, Reed-Solomon (RS), forward error correction (FEC) codes. These codes correct the bit errors after they occur, and

21、 must be sufficiently long to correct relatively long bursts of errors caused by impulse noise. Reed-Solomon codes also enhance performance with crosstalk and background noise by a few dB. For ADSL/2/2+ and VDSL2, error-causing impulse noise will usually wipe out an entire 250 microsecond DMT symbol

22、, or multiple consecutive DMT symbols. The INP value is the number of consecutive DMT symbols which can be reliably corrected. INP increases as delay increases and as coding redundancy increases. ADSL, ADSL2, and ADSL2+, originally only had INP = 1, but this has been extended to INP = 2 as an option

23、 for ADSL2+. VDSL2 supports INP as high as 16. The RS code block length is n bytes, the number of data bytes it carries is k, and the number of byte errors it can correct is t = (n - k)/2. Typically n = 255, and t = 8. Interleaving is used to extend the amount of correctable errors in a burst, with

24、interleaving depth D, a burst error of length D*t bytes may be corrected. Relations between coding and interleaving parameters for DSL are described in reference 2, and some specific values are in Table 7. Typically the resulting delay is no greater than 8 milliseconds, which is about equal to the l

25、ength of an entire interleaved and coded block. Uncorrectable errors generally result in this entire 8 milliseconds of data being lost, although many of the bits may still be correct. 21 ADSL, INP = 1 ADSL2+, INP = 2 Number of correctable DMT symbol errors, INP 1 2 Total bytes per RS codeword, N 255

26、 79 Correctable Byte Errors per RS codeword, T 8 5 Data bytes per codeword, K 239 69 Net data rate (Mbps) 7.5 24 Line data rate (Mbps) 8.0 27.48 Interleaver depth, D 32 352 Delay (Bytes) 7874 27378 Delay (millisec) 7.87 7.97 Span (millisec) 8.16 8.10 Correctable error burst length (bytes) 256 1760 C

27、orrectable error burst length (microsec) 256 512 # Codewords per DMT symbol = 1/S 1 11 Table 7: Some Representative Impulse Noise Error-Correction Parameters for DSL Models for the length of burst errors caused by impulse noise impulse noise have been presented 145. A received impulse can cause an e

28、rror if it is above roughly 1 to 5 milliVolts on a line without much excess SNR margin. It was found that the maximum magnitude of an impulse on a copper loop is closely modeled by the following hyperbolic distribution. The probability that the maximum magnitude of an impulse, H, is greater than h,

29、given that it is above some positive threshold, T is: ()()./,1|Pr2hThTThTHhH The error burst length B is the time that the impulse is above a certain error-causing threshold, T. The following model for burst error lengths caused by impulses was found to be reasonably accurate: ()2|Pr+=TbTTHbB(1) Mod

30、els were found using measured impulse noise data with T = Tj =0.947 milliVolts and = 33.6 microseconds 1. ADSL, ADSL2, ADSL2+, VDSL, and VDSL2 use discrete multi-tone (DMT) modulation and typically transmit a distinct DMT symbol once every 250 microseconds (4 kHz rate). If impulse noise causes an un

31、correctable error, then typically one entire DMT symbol is lost, but occasionally two or sometimes even three DMT symbols may be lost. Looking at equation (1), it is seen that = 31.8 is much less than 250 microseconds. So, a good approximation for the conditional distribution is that: 22 ()2|PrbxxBb

32、B And, the probability that a burst error length is between N and N DMT symbols, given that it is at least length N is: ()22111Pr+=NNNsymbolsDMTerroredofNumber (2) This approximation is shown to be reasonably accurate in Table 8, which also shows great inaccuracy in a nave geometric model such as th

33、at used in the Gilbert model of the previous section. Burst error length in Number 250 microsecond DMT symbols Exact Hyperbolic model probability (1) Approximate Hyperbolic probability (2) Geometric model 1 0.719 0.750 0.719 2 0.151 0.139 0.2023 0.055 0.049 0.1454 0.026 0.023 0.104 Table 8: DSL Burs

34、t Error Length Models The model estimates that 1.3% of impulses with maximum magnitude greater than 0.947 milliVolts are at least as long as a 250 microsecond DMT symbol. This agrees with measurements 5 reporting that impulses longer than 200 microseconds occurred 1% to 3% of the time. Assuming the

35、burst length model in equation (1), the number of DMT symbols that are wiped out by an impulse is estimated. Many variables affect the probability that the DMT symbol is in error: loop length, bit rate, impulse energy spectrum, etc. Reference 4 related these all to the energy in the impulse. Here, i

36、t is assumed that, only the length of the impulse determines how many DMT symbols are errored, assuming that at least one DMT symbol is errored. The length of a DMT symbol is 250 microseconds. It is assumed that a DMT symbol is errored if and only if at least 50 microseconds of the impulse occur in

37、the symbol. So, the minimum impulse burst length to make two DMT symbols errored is 100 microseconds, and only if exactly 50 microseconds of this impulse are in each DMT symbol are they both errored. Given that there is an error causing impulse, its arrival time is assumed to be uniformly distribute

38、d in a DMT symbol. The length of an impulse is modeled as in equation (1), and the arrival time is assumed uniformly distributed in a DMT symbol time. Arrival time and burst length are assumed to be independent, so the joint probability density is the product of the two densities. The conditional pr

39、obability of a burst length given that it is greater than 50 microseconds is found, and this multiplied by the uniform probability that the burst overlaps more than 50 23 microseconds with each number of DMT symbols, and these are summed for all possible burst lengths. The results are shown in Table

40、 9. # Errored DMT Symbols, N Pr(# errored DMT Symbols = N | at least one error) 1 0.848 2 0.1233 0.0174 0.0065 0.003 Table 9: Conditional Probability Estimate Of The Number Of Errored DMT Symbols Assuming That There Is At Least One Errored DMT Symbol, Before Error Correction The model is approximate

41、; it represents the aggregate performance of many lines and not that of any individual line. It shows that in about 85% of error cases 1 DMT symbol is errored, in about 12% of error cases 2 DMT symbols are errored, and in about 3% of error cases 3 or more DMT symbols are errored. Recall that only ab

42、out 2% of impulses cause any errors in the first place, on a line with little excess SNR margin. The arrival rate of impulses with maximum magnitude of 3 milliVolts or greater was approximated as one every 15 seconds 4, using measurements on several lines. So, with INP=1 DMT symbol protected from er

43、rors the uncorrectable error rate is roughly one every 1.4 hours, and with INP=2 DMT symbols protected from errors the uncorrectable error rate is roughly one every 7 hours. However, shorter loops often have fewer or no errors. Overall, increasing INP from 1 to 2 decreases the uncorrectable impulse

44、noise error rate by about a factor of 5. The impulses measured here on bare copper lines here may be of shorter duration than some additional impulse mechanisms. Intermittent bad connectors on many CPE may generally be interpreted as impulse noise when the line starts “jittering.“ This can cause the

45、 same effects as slightly bad splice, looking like an impulse periodically due to minute gyrations of the line with temperature, atmospheric pressure, etc. Also, POTS transients such as ring trip may leak through a splitter, causing long duration impulse errors. These may call for more frequent use

46、of higher values of INP. More extensive field data on impulse burst lengths could be used to help hone the mathematical model here. If there is an unrecoverable error, then an entire interleaved block is assumed lost; typically about eight milliseconds of data. ANNEX A: REFERENCES 1 K. Kerpez. Models for Packet Loss Bursts Related to Video Quality of Experience (QoE,) DSL Forum contribution dsl.2005.460.03, September 20-22, 2005. 2 C. Modlin. Proposed Restrictions on Interleaver Complexity, ATIS Standards Contribution T1E1.4/2003-493, May 24, 2004. 24