Per-segment based Full Passive Measurement of QoS for the FMC Environment Norihiro FUKUMOTO, Satoshi UEMURA, Hideaki YAMADA, Hajime NAKAMURA KDDI R&D Laboratories Inc. {fukumoto, sa-uemura, hd-yamada, nakamura}@kddilabs.jp Abstract Real-time multimedia applications apply not only to fixed IP networks, but also to the FMC (Fixed Mobile Convergence) environment composed of various infrastructures, their behavior of IP packet transmission differing significantly from each other. The purpose of this study is to present a novel service control scheme considering the per-segment based QoS (Quality of Service). The scheme also emphasizes backward compatibility with existing applications. y employing full passive monitoring of RTP/RTCP packets and modifying signaling messages, the scheme allows compatible service control. As an experimental result, the per-segment based measured QoS shows good agreement with simulation results. We suggest that our proposed scheme works sufficiently well for flexible service control. 1. Introduction In recent years, IP communications are expanding widely to a variety of infrastructures. As IP networks become more ubiquitous, FMC (Fixed Mobile Convergence), which is a seamless integration of wireless networks and fixed networks, has gained considerable attention. Research into real-time multimedia applications over IP networks has also been a frequent topic of debate. These real-time multimedia applications apply not only to fixed IP networks, but also to the FMC environment consisting of various infrastructures. It is well known that the behavior of IP packet transmission over wireless networks differs from that of fixed networks. Moreover, each wireless network has different characteristics depending on the infrastructure type, WiFi, WiMAX, 1x EV-DO, and others. Generally, wireless links provide high bit error rates depending on the radio link conditions and the congestion resulting in IP packet losses. In addition, the handover mechanism also causes large jitter and packet loss. Therefore, real-time multimedia applications should observe the conditions of each wireless network for suitable admission control and adjustment. For instance, limiting calls from a client in a bad sector of a wireless network in which transmission quality is poor can maintain the total user-level QoE (Quality of Experience). QoE is defined as the overall acceptability of an application or service, which is perceived subjectively by the end user or estimated objectively from the total transmission quality as QoS. Many real-time multimedia applications adopt RTP (Realtime Transport Protocol) protocol, which includes a QoS/QoE feedback function using RTCP (RTP Control Protocol) [1]. RTCP SR/RR (Sender Reports/Receiver Reports) and RTCP XR (extended Reports) [2] are used to feedback IP network conditions from RTP receivers to RTP senders. However, the original RTCP provides overall feedback on the quality of end-to-end networks as a whole, rather than each different network individually. The feedback provided on end-to-end network quality is insufficient for applications in the FMC environment, since the quality of end-to-end networks depends on the quality of each network. It is difficult to separate the quality of individual networks from that of overall end-toend networks. Furthermore, most clients do not consider service control with the feedback of RTCP SR. To solve this problem, a number of strategies have been developed in recent years[3]-[5]. However, they have excluded backward compatibility with existing applications. The purpose of this study is to present a novel persegment based full passive measurement of QoS and estimation of QoE for service control over the FMC environment by employing passive monitoring of RTP/RTCP packets. The scheme allows compatible service control with existing applications. In Section 2, we recall the network configuration. Section 3 and 4 deal with our proposed scheme for a full passive measurement of QoS and estimation of QoE. In Section 5, prototype implementation and experiments are discussed. Finally, we discuss the advantages of our proposal. 2. Requirements for real-time applications in the FMC environment We assume the FMC environment consists of different networks that are connected with a SC (Session order Controller) [6], which can certainly monitor and modify the SIP signaling messages[7] and RTP/RTCP streams pass through. We consider per-segment based QoS, which means the QoS of the network segments divided by SCs, an effective means of service control. Figure 1 shows a network configuration consisting of a mobile
network and a fixed network, their characteristics differing significantly from each other. We also suppose that the terminals send RTCP packets to measure the overall QoS/QoE end-to-end. The end-toend QoS/QoE notification is helpful when it is used over homogeneous networks, but it does not always work well over the supposed network topology. 3. Per-segment based QoS measurement schemes There are two approaches to obtaining QoS of each segment, active measurement and passive measurement. The active measurement scheme sends prove packets to measure QoS practically. The active scheme has advantages in improving the accuracy of the measured QoS, but the prove packets may raise severe unnecessary congestion, particularly in wireless networks. Therefore, we focus on passive measurement schemes by monitoring existing RTP streams. Several passive measurement schemes have been proposed to provide detailed per-segment based information on QoS deterioration as shown in Table 1. For instance, ITU-T recommendation P.564[8] presents a location correlation scenario that identifies the location of QoS deterioration in a segment by monitoring the quality in the same stream at multiple points. The per-segment based measured QoS can be applied to improve the resolution in dynamic transport resource management. All the listed schemes suppose the use of RTP and RTCP protocol. These reporting schemes consist of clients and network nodes. Network nodes are generally located on network junctions to measure and relay the per-segment based transmission quality. Network nodes can also modify RTP media streams and handle signaling protocols depending on the location. The most primitive scheme is an RTP-level translator and mixer oriented scheme. RTP translators and mixers are defined in [1] and detailed in Internet-Draft RTCP HR[3]. Translators and mixers possess similar intermediate functions that forward or combine RTP streams and may change the encoding of the RTP streams. esides forwarding or combining streaming packets, these also process RTCP packets that may contain the Figure 1 Network configuration per-segment based QoS. SubRTCP is an extension of RTCP which is suggested in Internet-Draft[4]. SubRTCP extends RTCP to multiple network nodes to allow monitoring of QoS and data delivery between clients and network nodes. This scheme is very similar to the RTCP HR scheme except for extensions of packet types that require particular clients. A Multi-RTCP scheme[5] has also been proposed for reporting per-segment based measured QoS. The Multi- RTCP Scheme informs clients about the per-segment based QoS using extensions of RTCP that require particular clients. A feature of this scheme is handling of RTP streams on network nodes. Network nodes do not terminate RTP streams, but only listen to them. This means that a network node failure will never infect the RTP stream. The aforementioned schemes are not perfectly compatible with existing basic clients because they are listen-only to RTP streams but they have reliance on the extension of RTCP. On the other, our proposed scheme is based on full passive QoS measurement, which means listen-only to not only RTP streams but also RTCP reports. This scheme depends on SCs which can handle RTP streams and signaling protocols. In this scheme, RTCP packets are sent only between clients, not SCs. SCs act as a border gateway and signaling gateway. They therefore never process RTP and RTCP packets, due to the fact that they are compatible with existing clients. The full passive QoS measurement scheme therefore provides network node-oriented control such as modifications to signaling protocols. On the other hand, the QoSs of each segment are not notified to clients, and they could not control RTP streams using it. In addition to Table 1 Schemes for per-segment based QoS measurement and QoE estimation RTCP HR SubRTCP Multi-RTCP Proposed Scheme The scheme is compatible with existing basic clients or not. asic clients can be used. Clients must be extended. Clients must be extended. asic clients can be used. Network nodes can modify the signaling packets or not. Signaling packets may not pass through. Signaling packets may not pass through. Signaling packets may not pass through. The SC can modify signaling packets that pass through. Clients can know the persegment based QoS or not. Clients may know persegment based QoS. Clients may know persegment based QoS. Clients may know persegment based QoS. Clients knows only end-to-end QoS. RTP handling on network nodes Terminated Terminated Pass through Terminated
that, the full passive mechanism can divide the network into only two portions, both sides of the SC, and hence the maximum number of divided segments for measuring QoS is limited to two segments. The information sharing mechanism between SCs is required to avoid the limitation. 4. Per-segment based full passive measurement of QoS for the FMC environment 4.1. Per-segment based full passive QoS measurement Figure 2 shows the structure of the per-segment based full passive QoS measurement scheme. The scheme is implemented on an SC; therefore, it can monitor RTP streams passing through it and measure their transmission quality. In addition to quality measurement of RTP streams, the scheme monitors corresponding RTCP packets including the quality of the whole network, and calculating their difference gives the per-segment based QoS (Figure 3). Detailed calculation of the quality of segment using the difference will be mentioned later. Moreover, the scheme can monitor SIP signaling packets passing through the SC and modify them for service control. To give an example of service control, it is conceivable that the SC removes the codecs for wideband networks to avoid excessive congestion. Figure 4 depicts an example of codec limitation by modifying the SIP signaling packet when the bandwidth is insufficient for wideband codecs. The SC removes the wideband codecs represented as "m=" and "a=" fields in the SIP signaling packet to prohibit them. The behavior of the full passive QoS measurement scheme depends on monitoring and modifying existing packets without any additional communication. Therefore, the scheme has good compatibility with existing clients and network nodes. 4.2. Measured QoS and estimated QoE parameters Some of QoS parameters are measured and contained in the RTCP packets. The extension of RTCP contains not only QoS parameters but also QoE parameters estimated from QoS parameters objectively. In this section, we describe how to measure and estimate the parameters of each segment. 4.2.1. Packet loss rate RTCP packets include packet loss rate information as the fraction lost in SR and the loss rate in XR. The following shows the process of calculating the packet loss rate using existing fields in RTCP packets. The SC can compute the number of packet losses in segment A ( N A ) counting the leap in the sequence number of RTP packets. The number of expected packets ( N ) and the packet loss rate of segment A ( LA ) are also calculated at the same time. Comparison of the cumulative number of packet loss rate in the current SR packet and the previous one gives the number of packet losses in the whole of segments A and ( N A ). Suppose the packet loss rate in segment is as follows: L = N /( N N ) A (1) When the previous SR packet is lost in segment, the SC can monitor packet loss rate information RTCP, which includes the packet loss rate of the whole of segments A and ( L A ). In this case, the packet loss rate in segment is represented as: L Figure 2 Per-segment based full passive QoS measurement scheme Figure 3 Structure of per-segment based full passive QoS measurement scheme Figure 4 An example of SIP signal modification based on the full passive QoS measurement scheme = L L )/(1 L ) (2) ( A A A
4.2.2. Round trip time Figure 5 depicts the calculation of the round trip time at the SC, where C is the captured time of the RTCP packets at the SC. The RTCP SR contains its own send time of itself T in the NTP timestamp field. The receiver of the RTCP SR returns the NTP time stamp in the LSR (Last SR timestamp) field in the returned RTCP SR. The SC can find the pair of RTCP SR using these fields. The round trip time of segment can then be obtained as: RTT C2 C DLSR (3) = 1 Client A T1 T2 RTCP SR (NTP timestamp = T1) RTCP SR (NTP timestamp = T2) RTCP SR (LSR = T1) Segment A C1 C2 SC Segment Client Figure 5 Round trip time calculations C 23 urst Find matched SR by NTP timestamp and LSR Gap C 31 C 14 4 DLSR (Delay since Last SR) 4.2.3. urst metrics RTCP XR also contains burst metrics fields that notify the statistics of the packet loss pattern[9]. The burst metrics report blocks are useful for identifying the cause of packet loss taking the features of the infrastructures into consideration. In the case of packet loss on a wireless link, there is a high probability that the packet loss is caused by a bit error in the infrastructure or packet drop due to congestion. They both cause packet loss, but they have a different packet loss pattern. Accordingly, the best adaptation should be selected depending on the burst metrics. In the case of congestion, new calls should be prohibited or reduced. It is essential to look at the persegment based packet loss patterns for estimating an accurate QoE and performing proper admission control. The burst metrics classifies the received RTP packet into two periods, burst and gap. A burst is a period during which a high proportion of packets are either lost or discarded. A gap is a period of low packet losses and/or discards. The burst metrics classify the packet into two periods standing on a four-state Markov model (Figure 6). The number of state transitions leads to the following parameters where C mn is the count of transition from state m to state n and P is the probability of the mn transition to state n when the current state is m. P 23 = 1 C22 /( C22 + C23) (4) P 32 = C32 /( C31 + C32 + C33) (5) burst density = P 23 /( P23 + P32 ) (6) gapdensity= C 14 /( C11 + C14 ) (7) The full passive QoS measurement scheme is applicable to RTCP XR for measuring the detailed cause of packet loss. Figure 7 shows the calculation sequence of burst metrics using the comparison of packet loss patterns. The SC monitors the RTP packet stream passing through to obtain the packet loss pattern firstly in segment A. The receiver of the RTP stream feeds back the packet loss pattern of the whole network as the Loss RLE Report C 22 2 3 1 C 32 C 13 C 41 C 33 C 11 state 1: received a packet during a gap state 2: received a packet during a burst state 3: lost a packet during a burst state 4: lost an isolated packet during a gap Figure 6 Four-state Markov model Figure 7 urst metrics calculation sequence lock of RTCP XR. The SC also monitors the packet loss pattern of the whole network and compares it with the packet loss pattern of segment A to calculate the proper packet loss pattern of segment. The burst metrics parameters can be calculated in a way similar to that for measuring the burst characteristics by clients. 4.2.4. R value and mean opinion score Generally, the MOS is estimated using ITU-T recommendation G.107[10], which defines the E-model, a computational model for use in transmission planning. The E-model offers not only a planning model but also the objective speech quality in the R value, which provides scoring within a range from 0 to 100 in the
narrow-band case where higher scores indicate better quality. The R value predicts the perceptive effects of different types of degradation on overall speech communication quality and correlates closely with the quality of the user experience. The computed R value can be mapped to the MOS. To compute the R value using the E-model, the aforementioned packet loss rate and round trip time are used. A method employing speech quality estimation with burst metrics has also been examined[11]. Consequently, it is possible to estimate the MOS using the known QoS parameters. 5. Experiments To verify our scheme, we developed a prototype system. In this section, we describe this system and show the experimental results. 5.1. Prototype implementations Figure 8 shows the prototype system of the full passive QoS measurement-enabled SC and the RTCP XRenabled cellular phone. The prototype was implemented based on the IETF specification of RTCP SR and XR. The VoIP application implemented on the cellular phone also supports SIP/SDP based signaling, which should be rewritten by the SC according to the per-segment based quality. 5.2. Mean opinion score The proposed SC was also implemented using the full passive QoS measurement scheme. Some experiments were made to verify the efficiency of the proposal. RTP clients are connected by an Ethernet LAN, which is divided into two independent segments by a full passive QoS measurement scheme-enabled SC. The divided networks are supposed to be the FMC environment composed of different infrastructures, fixed and mobile networks. RTP/RTCP streams pass through these segments successively. Each segment is simulated by a network simulator which provides the same amount of random packet loss for both segments. Figure 9 shows the result of QoE estimation on the SC. The 1st or 2nd segment means the segment through which the RTP streams passed first or after the SC. The mean opinion scores are mostly in agreement with the calculated values in both segments. 5.3. urst Metrics In the experiments for burst metrics, the Gilbert-Elliott channel model defined in ITU-T Recommendation Figure 8 Full passive estimation-enabled SC and RTCP XR-enabled cellular phone mean opinion score Estimated QoE parameters 4.5 4.0 3.5 Measured QoS parameters 3.0 0.0 1.0 2.0 3.0 4.0 5.0 packet loss rate (%) G.191[12] was used as the packet loss model. The model is a state machine which has two states, good and bad. The packet loss rates in the good state and the bad state assume that P G = 0. 0 and P = 0.5. The bursty nature of the model is given by a variable γ ( 0.0 γ < 1.0) with a higher value indicating more bursty packet loss. The transition probabilities are set as follows where f is the packet loss rate. ( 1 ( P f ) /( P P )) simulation experiment (1st segment) experiment (2nd segment) Figure 9 Experimental results of mean opinion score p = ( 1 γ ) G (8) q = 1 p γ (9) γ was set to 0.2 for random packet loss, and to 0.8 for burst packet loss. We employed a network consisting of two segments for the experiments, a combination of random and burst packet loss. Separation of packet loss characteristics was performed for verifying the accuracy of the full passive QoS measurement scheme. Figures 10 show the experimental results and comparisons with the simulation result locally generated. Simulation results were obtained with random packet loss, burst packet loss, tandem connection of random and burst packet loss, and their reverse order. The tandem connection model indicates the
medium metrics, which can be measured without the full passive measurement scheme. The graphs also include the per-segment based burst metrics measured by the full passive QoS measurement scheme. The experimental and simulated results are in close agreement; they reveal that the scheme separates the burstiness of the two different segments exactly. Consequently, the SC can handle the network condition of each segment and perform suitable admission control. 6. Conclusions In this paper, we presented a full passive QoS measurement scheme and QoE estimation scheme, which is a simple framework to measure the per-segment based QoS. The scheme satisfies backward compatibility by employing full passive monitoring of RTP/RTCP packets and modifying signaling messages. We presented the detailed calculation method of the R value, mean opinion score, packet loss rate, round trip time, and burst metrics. We also gave an example of service control to avoid excessive congestion. The experimental results on mean opinion score and burst metrics for each segment demonstrated good agreement between the values measured by SC and calculated values using simulation. Among the burst metrics tested, the most efficient metrics was found to be burst density and gap density. From the results, we can conclude that the full passive QoS measurement scheme separates the QoS of individual segments from that of overall end-to-end networks. The results obtained show that the full passive QoS measurement scheme gives relatively accurate results and is sufficient for service control. The SC can handle the per-segment based QoS and give suitable admission control using them. Acknowledgements The authors would like to express their gratitude to Dr. M. Suzuki and Dr. S. Akiba of KDDI R&D Laboratories, Inc. for their support of our work. This work was partly supported by the National Institute of Information and Communications Technology (NICT). References [1] H. Schulzrinne, S. Casner, R. Frederick and V. Jacobson, RTP: A Transport Protocol for Real-Time Applications, IETF RFC 3550, July 2003. [2] T. Friedman, R. Caceres, and A. Clark, RTP Control Protocol Extended Reports (RTCP XR), IETF RFC 3611, November 2003. [3] A. Clark, A. Pendleton, R. Kumar, and K. Connor, RTCP HR - High Resolution VoIP Metrics Report locks, IETF Internet-Draft, burst density (%) gap density (%) 100 90 80 70 60 50 40 30 20 10 simulation (random) simulation (random+burst) experiment (random as 1st segment) experiment (burst as 1st segment) 0 0.0 1.0 2.0 3.0 4.0 5.0 2.0 1.8 1.6 1.4 1.2 1.0 0.8 0.6 0.4 0.2 simulation (random) simulation (random+burst) experiment (random as 1st segment) experiment (burst as 1st segment) packet loss rate (%) simulation (burst) simulation (burst+random) experiment (random as 2nd segment) experiment (burst as 2nd segment) 0.0 0.0 1.0 2.0 3.0 4.0 5.0 packet loss rate (%) simulation (burst) simulation (burst+random) experiment (random as 2nd segment) experiment (burst as 2nd segment) Figure 10 Experimental results of burst/gap http://www.ietf.org/internet-drafts/draft-ietf-avt-rtcphr-00.txt, February 2007. [4] X. Xu, K. Connor, T. Shaikh, R. Padmanabhan, and I. Umansky, SubRTCP:RTCP Extension for Internal Monitoring of RTP Sessions, IETF Internet-Draft http://www.ietf.org/internet-drafts/draft-xu-avtsubrtcp-00.txt, December 2006. [5] N. Fukumoto, H. Yamada, H. Kawai, Evaluation Result of Transmission Control Mechanism for Multimedia Streams based on the Multi-RTCP Scheme over Multiple IP-ased Networks, CCNC 2006, MP1-06-4, January 2006. [6] J. Hautakorpi, G. Camarillo, M. hatia, R. F. Penfield, and A. Hawrylyshen, Requirements from SIP (Session Initiation Protocol) Session order Control Deployments, IETF Internet-Draft, http://www.ietf.org/internet-drafts/draft-ietf-sipping-sbc-funcs-01.txt, February 2007. [7] H. Schulzrinne, G. Camarillo, A. Johnston, J. Peterson, R. Sparks, M. Handley, and E. Schooler, SIP: Session Initiation Protocol, IETF RFC 3261, June 2002. [8] ITU-T Recommendation P.564, Conformance testing for narrowband voice over IP transmission quality assessment models, July 2006. [9] A. D. Clark, Modeling the Effects of urst Packet Loss and Recency on Subjective Voice Quality, Proceedings of IP Telephony Workshop 2001, pp.123-127, April 2001. [10] ITU-T Recommendation G.107, The E-model, a computational model for use in transmission planning, International Telecommunication Union, May 2000. [11] S. Uemura, N. Fukumoto, H. Yamada, and H. Nakamura, A study of QoS/QoE measurement system implemented on cellular phone for NGN, Proceedings of the 2007 IEICE Society Conference, S-10-7, September 2007. [12] ITU-T Recommendation G.191, Software tools for speech and audio coding standardization, International Telecommunication Union, November 2000.