NEW_NETWORK_QOS_MEASURES_FOR_FEC-BASED_AUDIO_APPLICATIONS_ON_THE_INTERNET.doc

此文档收集于网络，如有侵权，请联系网站删除NEW NETWORK QOS MEASURES FOR FEC-BASED AUDIO APPLICATIONS ON THE INTERNETTeruko MIYATA Harumoto FUKUDA Satoshi ONONTT Software LaboratoriesAbstractNew network QoS(Quality of Service) measures for interactive audio applications using FEC(Forward Error Control) are proposed. Applications such as internet phone require both low data loss and a short delay for the underlying transport layer. FEC-based error control has become popular as a way of metting such requiremenets; new application data (or highly compressed data) are copied onto successive packets, so that random packet losses in networks can be concealed to some extent. From the viewpoint of FEC-based applications, actual QoS depends not only on the loss and delay of each packet. Conventional network QoS measures such as loss rate and delay distribution, however, only focus on each packet. Therefore, the probability of long successive losses, for example, cannot be monitored, even though they strongly affect FEC-based application QoS. We propose a new concept named “Loss Window Size” for measuring the QoS of successive packets. Definitions of loss and delay are generalized using this concept. These definitions take FEC-based error concealment into account. Therefore, these measures enable more precise estimation of FEC-based application-level QoS than conventional measures. In order to show the effectiveness of the proposed measures, we have built an experimental monitoring system on working networks. The actual data show that network QoS may vary from time to time in terms of newly defined measures, even though QoS variation using conventional measures are not so apparent. We also model FEC-based application QoS, and show that application QoS and proposed network measures correspond well.1.IntroductionApplications using continuous stream, such as audio and video, are becoming popular over the Internet. Interactive audio applications such as an Internet phone, require low loss and a short playout delay. Voice data loss causes voice cracking, and a long delay spoils interactivity. Therefore, methods for controlling packet loss in networks and error control methods l for concealing packct loss have been developed for continuous stream applications. The packet loss control methods reserve resources along thc communication paths. For example, there exist bandwidth reservation methods such as RSVP and ST-I1 2, 3. Comniunications using reserved paths are expected to be loss-free. Error control methods, in contrast, tolerate packet losses and simply try to conceal them. Error control methods are categorized as either Automatic Repeat reQuest (ARQ) or Forward Error Correction (FEC) 4. An ARQ method automatically retransmits errored or lost packets, and make such errors invisible to applications. One popular example is the TCP protocol. However, repeated retransmission may cause significant delay ancl jitter when the transmission quality is poor. For this reason, the ARQ method is not suitable for interactive audio applications, which need a real-time response. An FEC method, instead of resending lost packets, sends enough redundant information for application data so that the data stream can be reconstructed even if some parts are lost. For example, multiple copies of data may be sent on N successive packets. If any of these are received, the application data can be constructed. FEC methods keep the data loss rate low when the redundancy factor is suitably selected. Furthermore, because it does not depend on retransmission, delay and jitter will be smaller than the delay and jitter of ARQ methods. Thus, for real-time audio applications using voice, FEC is better than ARQ. FEC methods, however, add redundancy to communication data, thus lengthening transmission times. The optimal redundancy factor depends on ensuring a real-time response, minimizing the delay, and also the network quality. With FEC metllods, a pattern of successive packet losses is iinportant. However, current expressions for network quality, such as loss and delay, are based on one packet oiily, ignoring loss patterns. Therefore, the probability of long successive losses, for example, cannot be monitored, even though they strongly affect FEC-based application QoS. In this paper, we propose a new method of expressing network and audio qualities for FEC-basecl audio conimnnication. Firstly, we begin by describing an error coiitrol nietjliotl suitalde for voice t,ransmission. Next, we define an “LW”called “LossWindowSize”that extends the definitions for packet loss and delay to groups of packets. Then we present information on packet loss and packet delay nieasured on the Internet, and discuss the results from t,he viewpoint of packet, loss and delay based on LW. Finally, we show the effectiveness of tthe proposed iiieasiires. The actaid dat,a show that, network QoS may vary from time to time in terms of newly defiiied measures, even though QoS variations using conventional measures are iiot, so apparent. We also model FEC-based application QoS, and show that app1icat)ion QoS and proposed network measures correspond well.2 Expression of network quality based on LW In this section, new measiires expressing iiet,work and application QoS using au FEC method are presented. 2.1 The need for expression of network quality based on LW(N) We assiiiiie that newly arrivecl data (or highly compressed data) di are copied on N successive packets as shown in Fig. 1. We treat these packets as one group, so we extend the definition of t,he coiiventional expression of network quality, such as loss and delay d, based on one packet. To handle duplicat,ed data on N successive packets, “LossWindow Size”is introduced. Intiiitively speaking, packetss are analyzed though a “LossWindow” of size N (packets). The window moves one by one, like a computing moving average. We write LW for LossWindow Size and LW(N) means “LW=N”. In the following, we extend loss and delay using LW.Network quality is conventionally based on one packet. Using LW, tjhis sitnation corresponds to LW( 1). Under our assumption above that successive packets carry the same data, this definition is not suitable. For example, when k successive packets are lost, the packet loss is equal to k in the conventional definition. However, in the case of LW(N), for example, if k d (data too late).2.3 Calculation of loss rate based on LW(N) To calculate all loss rates LW(N)(1Nn), as a direct method, we calculate the number of N successive packet losses for each LW(N). For all received packets (total: T), this method needs to scan all received packets n times. Thus, the cost is O(T*n) to compute the LossRate N for every N (1Nn). We propose a faster method in which we can reduce the calculation cost. First of all, we define “losses with run length: k”; when a packet pi reaches a destination, packets pi+1, pi+2, pi+k are lost, and packet pi+k+1 reaches the same destination, then we define such a situation as “losses with run length: k”. ck: = the count of losses with run-length k. We assume K to be the longest loss run-length, and T to be the total number of received packets. As shown in Fig. 5, in a pattern of losses with run-length = k (1 k K), LW(N)-L occurs. Each number of LW(N)-L occurrences in this pattern follows (we denote a number of occurrences for LW(N)-L as “|LW(N) L|”):The packet losses with run-length k occur ck times,so the number of occurrences of LW(N)-L (we denote this as “LNk” for each k is given by the following: Then the total number of LW(N)-L(l N n) “LN” isBased on the above, we scan once for all T packets to calculate ck; then, we can calculate the LossRateN for any N (1 N n). Consequently, using this method, this cost is O(T + nK) for computing Loss-RateN for every N (1 N n). In general, the longest, loss run-lengt K T, and n T, then T + nK Tn. Thus, we can reduce the number of scans.For example, consider the packet-loss pattern in Fig. 6. In this pattern, c1=2, c2=1, and c3=1. So each LossRatel, LossRate 2, and LossRate 3 under LW(1), LW(2), and LW(3) is:3 Experimental monitoring networkIn order to show the effectiveness of the proposed measures, we have made the monitoring system on a working network as shown in Fig. 7. We have moiiitored packets flowing from a stream generator HOST 3 at a network 2 to HOST 1 at the network 1. Packets are probed at network 2 using HOST 2, and also at network 2 using HOST 1. Both hosts were equipped with the QoS Visualizer 6, 7. Thus, each probed packet was given precise timestamp synchronized to UTC (Universal Coordinated Time) within 0.5 ms. HOST 2 also works as an NTP server, and HOST 3 is its client. Generated packets are transported using UDP, with the fields shown in Fig. 8. The sequence of numbers begins from 0, and is increniented by one when a packet is generated. The packet length is 320 octets, and the packets are generated at 40-ms intervals. By comparing these timestamps between HOST 1 and HOST 2, we can collect information on network quality such as one-way delay and jitter. Checking packets only observed at HOST 2, losses can also be detected. Thus we can also obtain packet loss patterns.4 Experiments results In this section, we present the network quality, applying LW(N) based on InterCase measrtrments.l LW(N)-L We used two hosts: HOST 1 is located in Tokyo, HOST 2 is outside Tokyo. We measured from 12:30 to 13:30. The networks (average) loss rate was 1.27%, and the longest, loss run-length was 27. The run-length distribution of the packet losses is shown in Fig. 9. Using this distribution and applying the calculation method described in Sec. 2, we obtained each LW(TJ)-L rate for 1 N 27 (Fig. 10). We also measured the loss rate for different days and times (Fig. 11). The measurement periods were all one hour and the loss rates for each network are shown below, where K means the longest loss run-length. The network with the largest K and loss rate was “Case A” and that with the lowest was “Case B”, where the loss rate was 0.81% (Table. 1).l LW(N)-D We observed each delay distributions on Case A and B.The distribution of LW(1)-D is shown in Fig. 12. The interval of the log-scaled histogram is determined so that there exist 100 points from any value c to 10 times c. For examining the results of LW(N)(2 N 4), we magnify the following 2 ranges (Figs. 13 and 14).(x: time of LW(N)-D, y: number of occurrences)l LW(N)-PE For LW(N)-PE(d) (d= 25, 50, 100, 200, and 400 ms). For Case A and B, we examined the probability of LW(N)-PE(d) in each 30-s interval. The results are as follows: 1. LW(N)-PE(d) of Case A (d=25, 50, 100, 200, and 400 ms)(Fig. 15). 2. LW(N)-PE(400 ms) of Case A (N=1,8) (Figs. 16 and 17) 3. LW(N)-PE(400 ms) of Case B (N=1,8) (Fig. 18 and 19)5 DiscussionSuccessive packet losses and LW(N)-L The run-length distribution of successive packet losses (Fig. 9) shows that single or two successive packet losses happen frequently, while three or more successive losses are rare. As shown in Fig. 10, LW(N)-L decreases when N increases. Even considering this, LW( 1)-L seems to be a singular point, since its probability is especially high compared with other LW(N)-Ls. Furthermore, as shown in Fig. 11, we have observed a partial linear logarithmic relation among LW(N)-L(2 N 8). Note that coefficients of this linear relation have changed considerably when observation periods were different. The conventional packet loss rate is an average rate for LW(1). Thus, it cannot express the difference of quality between Case A and B. From only the loss rate, the QoS difference of Case A (1.27%) and B (0.81%) is not so apparent. In contrast, LW(N)-L can express a significant difference in successive loss patterns. For example, Fig. 11 shows that in Case A, the loss rate under LW(9)-L is about 0.1%. In contrast, in Case B, the loss rate becomes 0.1% when we change to LW(4). LW(N)-D If reclundancy is provided by N successive packets having the same data, loss can be avoided, because if at least, one of the packets reaches the destination, the original packet can be reconstructed. While this redundancy reduces packet losses, it prolongs delay. Increasing the N of LW(N) increases the delay. However, as shown in Fig. 13, some groups (N- successive packets) are recovered from losses which are the loss case under the LW(1). Similarly, Fig. 14 expresses the effect of LW(3) or longer. LW(N)-PE(d) In our experimental system, a packet is generated and sent at regular 40-ms intervals. Tlierefore, in the case of LW(8), the minimum playout, delay is 280 ms, and encodiiig and decoding need some time. Considering the above mentioned conditions, we set the threshold for playout time d at 25, 50, 100, 200, and 400 ms. On Case A, as shown in Fig. 15, although we waited for the packet from 200 to 400 ms, LW(1)-PE(d) did not) improve. This means that playout errors are causecl mainly by packet lowes when the playout delay exceeds 200 ms.Application-level QoS model Audio applicatioiis such as an InterCse phone, in general, tolerate a few playout errors. To niodel FEC-based application-level QoS, we define “error condi- tion”as the duration where LW(N)-PE(d) exceeds a specified threshold. We use 30 s as measurement period, and use 0.5, 1.0, and 2.0% as thresholds. Application-level QoS is modeled by the ratio of the “error condition” time in total playout time. The results of the “error condition!ratio for each threshold are shown in Fig. 20 and 21. In these figures, tlie playout, delay is always set, to 400 ms. We assume user requirements to be no “error conditions” with threshold 2.0%. On Case B, it can be easily achieved by setting N to be 2. On Case A, however, even if tuning any N (1 N 8), keeping the error condition under 5% is difficult. Furthermore, in any threshold (0.5, 1.0, and LO%), it is apparent, that Case B provides better QoS tlian Case A, from the viewpoint of tlie error condition ratio. As shown in the previous section, LW(N)- generalized loss, delay, and playout error distinguishes the QoS of Case A and Case B quite well. Thus, the proposed measures can be said to express network QoS better than conventional measures, for the purpose of estiniating the QoS of FEC-based applications.6 Conclusion We proposed a new concept LW(N) for measuring the QoS of successive packets. We built an experimental monitoring system on working networks. An analysis under this system and its results showed that loss and delay based on LW(N) more precise enabled estimation of FEC-based application-level QoS than conventional QoS expressions. Thus we can observe network QoS from the viewpoint of FEC-based applications