Playout Control for Streaming Applications by Statistical Delay Analysis Kouhei Fujimoto Graduate School of Engineering Science Osaka University E-mail : k-fujimo@ics.es.osaka-u.ac.jp Thank you, Mr. Chairman. It’s my great pleasure to give a talk to you. I’ll be speaking about Playout Control for Streaming Applications by Statistical Delay Analysis

Contents Introduction The analysis of packet delay characteristics Issues on Streaming Applications in the Internet Playout buffer algorithm objectives The analysis of packet delay characteristics Measurement method Analytic approach to estimate parameters of distribution function Modeling method Application of the analytic results to real-time applications Proposed playout buffer algorithm Performance evaluation Implementation Conclusion The contents of my presentation are shown here. First of all, I’ll explain the issues on Streaming Applications in the Internet, playout buffer algorrithm and objectives of my research. Next, I’ll show how to analyze the packet delay characteristics. ( Measurement method, analytic approach to estimate parameters of distribution function and modeling method. ) Next, I’ll explain how to apply the analytic results to real-time applications. ( Here, I propose an adaptive playout buffer algorithm based on the results of statistical analysis. Then I’ll evaluate our algorthm, and actually implement the algorithm to the real-time application. And finally, I’ll conclude my presentation.

Real-time Applications server The applications which playback audio data as soon as download Interactive Voice, Voice conference （vat、NeVoT..） Radio （Winamp, RealPlayer..） etc ♪ ♪♪ Internet Source file ♪ Telephone line AP According to the fast growth of the Internet, there are an increasing number of network applications. The real-time audio is one of such applications. The use of real-time audio includes interactive voice, voice conference and internet radio, And these applications become widely used. client

Issues on Real-time Applications in the Internet 2 3 1 4 Jitter Delay Variation Packet Loss server client However, currently available, best effort service model was never engineered to handle real-time traffic, since it provides variable jitters, delays and loss ratio. So, compensating for jitters and variable delays, and detecting packet losses are accomplished primarily through playout buffer algorithm. Playout buffer algorithm

Playout Buffer Algorithm Jitter reorder the packet in the buffer Delay Variation absorbed by buffering Packet Loss detect packets which will not arrive Playout buffer algorithm sets up some buffers at the client, and stores the arrived packets in the buffer for some time. So, this algorithm can reorder the packets which arrived at the client, and absorb the delay variations. Moreover, it can address all packets which cannot arrive at client host in time as packet losses.

Playout Delay Playout delay = buffer size + delay Playout delay affects the quality of the real-time application small buffer size the client treats packets to be lost even if those packets eventually arrive in a short time real-time > high quality for interactive voice, voice.conference Large buffer size playout more packets, but cannot playout in real-time real-time < high quality for internet radio Playout buffer algorithm stores the packets in the buffer until the playback. Playout delay consists of this storing time, and packet transfer delay. In other word, playout delay means the deadline until playouting the audio data of each packet. So, playout delay directly affects the communication quality of the real-time application. For example, too short playout delay leads to the fact that the client treats packets to be lost even if those packets eventually arrive in a short time. As the results, the client can playout the packet in a short time from packet arrival, but the quality of audio becomes worse. (Of course, large playout delay may introduce an unacceptable delay, which can playout more packets, but cannot playout in real-time.) Thus, we need carefully determine the playout delay which satisfies the application’s characteristics. we need carefully determine the playout delay which satisfies the application’s characteristics

Delay Distribution and Target Value 220 240 260 280 300 320 Delay [ms] 10 8 4 6 [%] 90% 9% 1% 2 Interactive Voice, Audio Conference Radio Too bad quality Too large Delay (non real-time) tail part of distribution Therefore, I determine the precise playout delay from the analysis of packet delay by statistical method, To clear the packet delay characteristic, I measure the delay in the Internet, and model the delay distribution in the colored green area. The reason why I model the part of delay distribution is because the delays in the green area are applicable for the deadline until playouting the audio data of each packet. I call this green area “tail part of distribution”. And, I define “target value” as the ratio of packets which satisfy the application’s quality. For example, I think the target value of internet radio is about 99%. Target value = the ratio of packets which will be playouted required by application

Objectives Analyze the packet delay characteristics Model the tail part of delay distribution Apply the analytic results to the real-time application Propose a new playout buffer algorithm which can control the packet loss ratio Evaluate the performance of new algorithm Implementation test (The studies about the characteristics of the delay have been made by some researchers, but most of those studies have focused on the average characteristics and entire distribution only.) So, I analyze the packet delay characteristics by modeling the tail part of delay distribution, for playout buffer algorithm. Then, I apply the analytic results to the playout buffer algorithm. Furthermore, I evaluate the performance of new algorithm and implement it.

The Analysis of Packet Delay Characteristics From now, I’ll explain the analysis of packet delay characteristics.

Measurement Method of One-way Delay Developed a measurement tool the sender records the current time into the packet the delay is calculated by the receiver host Synchronize the sender and receiver hosts clocks Packet size : 160 byte Interval : 80 ms Internet 28.8Kbps Telephone line Sender host Receiver host Synchronize The clocks GPS satellite ISP LAN To measure the one-way delay, I developed the server-client based tool. In the tool, the sender host records the current time into the packet before sending. When the packet arrives at the receiver host , the delay is calculated by the receiver host. In this way, time clocks of the sender and receiver should be synchronized for proper delay measurement. So, We use GPS for time synchronization. In my experimental setting, ( the sender host accesses to the Internet via LAN, and the receiver host was connected by the 28.8 Kbps modems to several ISP.) The packet size was set to be one hundred sixty (160) bytes and an interval of packet emissions was fixed at 80 ms.

Modeling Method Pick up four distribution functions as candidates to adequately represent delay distribution Normal, LogNormal, Exponential, Pareto Select Adequate function as the model Estimate parameters of each functions from the delay at the tail part of distribution Determine the most appropriate function by χ-squared test in the tail part of distribution In this slide, I explain the method to model the packet delay distribution. At first, I select four distribution functions as candidates to adequately represent delay distribution; Normal, LogNormal, Exponential and Pareto distribution is the candidates. Secondly, I determine the most appropriate function as the model by chi-squared test. Here, in order to detect the packet loss from the distribution of packet delays, the coincidence at the tail part of distribution is very important. Thus, I estimate the parameters of each candidates respectively from the delay at the tail part of distribution. Furthermore, I determine the most appropriate function by chi-squared test in the tail part of distribution

Comparisons among Delay Distribution and Candidate Functions X-squared Normal 122.23 Exponential 471.57 LogNormal 103.56 Pareto 74.32 0.1 Exponential(471.57) 0.01 Probability Density Target Value(99%) Pareto(74.32) Next, I explain the results of the measurement and modeling. This figure compares the distribution of the measured One-way delay with four candidates functions in busy hours. We set the target value to 99% . The line labeled “delay” represents the tail part of the cumulative density distribution of the One-way delays. And the other lines represent the candidate functions. This table shows the chi-squared test results. In chi-squared test, the smaller the test result shows, the more appropriate the distribution is as the model. So, from this table,because the Pareto distribution has the smallest chi-squared value, I found that the Pareto distribution is the most appropriate as the model. Busy hour Measurement for 40 min 13,000 data Normal(122.23) Delay(Heavy) Log-Normal(103.56) 0.001 200 400 600 800 1000 1200 Delay [ms]

Χ-squared Test Results Hour Χ-squared Test Normal Exponential LogNormal Pareto 9PM 83.82 602.56 71.96 19.56 11PM 53.86 470.90 49.67 30.10 1AM 55.06 426.46 49.99 24.01 5AM 94.45 500.91 85.77 25.16 9AM 107.76 754.09 98.74 45.33 12PM 108.66 1218.95 101.09 30.61 3PM 109.07 336.49 85.41 21.21 Next, I summarize the results of chi-squared test in this table. The first column shows measured time and second through fifth columns show the results of the chi-squared test. We can observe that the Pareto distribution has always the smallest value in all experiments from this table . So, Pareto distribution is the most appropriate as the model regardless of time of day. The Pareto distribution is the most appropriate as the model regardless of time of day

Application of the Analytic Results to Real-time Applications We have found that we can model the tail part of delay distribution by the Pareto distribution. From now, I’ll show how to apply the Pareto distribution to real-time application.

Proposed Playout Buffer Algorithm Use the model to determine the adequate playout delay which satisfy the quality required by the applications Record the delays of last n packets Estimate parameters of Pareto distribution function from the distribution of recorded delays From estimated Pareto distribution, calculate the playout delay which keep the packet loss ratio to satisfy the target value I propose a new algorithm using the model to determine the adequate playout delay which satisfy the quality required by the applications. At first, the algorithm records the delays of the arrived packets. Secondly, the algorithm estimates parameters of Pareto distribution function from the distribution of recorded delays. At last, the algorithm calculates the playout delay which keep the packet loss ratio to satisfy the target value, from estimated Pareto distribution.

Simulation - Playout Delay Variation Test the performance of proposed algorithm by tracing measured packet delay 100 200 300 400 500 600 700 800 900 1000 16000 16200 16400 16600 16800 17000 Sequence Number Delay Algorithm1(1.40) Algorithm2(5.95) Algorithm3(0.05) Algorithm4(4.30) Playout Delay[ms] Target value = 99% (11PM) In my simulation, we test the performance of proposed algorithm by tracing measured packet delays. This figure compares the playout delay variations among four algorithms in busy hour. Algorithm 1, plotted by red line, is our proposed algorithm. Algorithm 2 through 4, is defined in paper [1]. The algorithm 2 and 3, plotted by green and pink lines, aim to keep no packet loss. The algorithm 4, plotted by blue line, aims to follow the delay variation. The target value is set at 99%. From this figure, we can find that the delay variation of algorithm 2 is underestimated, from large loss ratio. On the other hand, the algorithm 3 has too large playout delay, compared with the other algorithms. Although the playout delay of Algorithm 4 can follow the drastic variation of one-way delay, its loss ratio is higher compared with proposed algorithm. We can observe the loss ratio of proposed algorithm can satisfy the target value. [1] R.Ramjee, J.Kurose, D.Towsley, and H.Schulzrinne, “Adaptive playout mechanisms for packetized audio applications in wide-area networks,” in Proceedings of IEEE INFOCOM’94, pp.680-688, April 1994

Performance Evaluation Hour Algorithm Target Value[%] Packet Loss Ratio[%] Avg. Playout Delay[ms] 11PM 1 95 5.13 221.17 99 1.37 265.12 99.9 0.14 855.38 2 - 2.46 237.16 3 0.24 392.64 4 3.95 230.72 Next, we compare the packet loss ratio and the average of playout delay. In the case that the target value is 95%, the packet loss ratio of proposed algorithm results in 5.13%. When the target value is 99%, actual loss ratio is 1.37. and when the target value is 99.9, actual loss ratio is 0.14. The experiments based on the data measured in another time , also show the same results. These results means that our proposed algorithm can control the playout delay with satisfying the target loss probability. Proposed algorithm can control the playout delay with satisfying the target loss probability

Implementation Apply proposed playout buffer algorithm to the real-time application ( input plug-in of winamp ) Test the performance of the algorithm Target Value[%] Packet Loss Ratio[%] Recorded audio data 90 8.5 99 1.8 99.9 0.2 Finally, I apply the proposed playout buffer algorithm to the real-time application, Winamp. I eventually run the winamp including our proposed algorithm, and evaluate whether the algorithm can control the quality, in real-time. Implementation test shows the same results as the trace-driven simulation.

Conclusion From statistically analytic method, I have found that the Pareto distribution is most appropriate as the model of the one-way delay’s distribution I have proposed a playout buffer algorithm based on the analytic results in order to control the quality of the applications. Numerical examples have shown that the proposed algorithm can control the playout delay satisfying the target packet loss probability. I’d like to conclude as follows. First, I have found that the Pareto distribution is most appropriate as the model of delay distribution from statistically analytic method. Moreover, I have proposed an adaptive playout buffer algorithm base on my analysis. Numerically, analysis and implementation test have shown that proposed algorithm can control the playout delay with satisfying the target loss probability Thank you for your kind attention.

Simulation Test the performance of proposed algorithm by tracing measured packet delay Keep the required quality? compare the performance of proposed algorithm with those of well-known algorithms shown in paper[1] In my simulation, we test the performance of proposed algorithm by tracing measured packet delays. Specifically, I test whether the algorithm can keep the target value or not. To evaluate the performance, I compare the performance of proposed algorithm with the performance of well-known algorithms shown in paper one. [1] R.Ramjee, J.Kurose, D.Towsley, and H.Schulzrinne, “Adaptive playout mechanisms for packetized audio applications in wide-area networks,” in Proceedings of IEEE INFOCOM’94, pp.680-688, April 1994

Factor Affecting Modeling Delay Distribution Time of day Busy time heavy traffic intensive variation of delay, increase of the packet loss ratio Day time, midnight opposite to busy time characteristic Throughout the measurement and the modeling, I need investigate the influences of the time of day. It is known that the Internet traffic pattern changes by time of day. For example, in working hours, the traffic is comparatively heavy, the variation of the delay becomes intensive.

Required Number of Packets for Reliable Parameter Estimation No drastic variation by changing sample number 0.2 0.4 0.6 0.8 1 1.2 1.4 500 1000 1500 2000 2500 3000 3500 4000 Packet Loss Ratio [%] The Number of Samples Sample1(Oneway,2 PM) Sample2(Oneway,11PM) Sample3(Oneway,3 AM) Sample4(Oneway,9 AM) Nunber = 500 In proposed algorithm, the number of measurements for the parameter estimation becomes a dominant factor. The accuracy of parameter estimation can be increasing the number of measurement data. However, on the other hand, the larger number of sample prevents the algorithm from following the dynamic change of the network condition, and the playout delay control cannot follow the drastic variation of one-way delay. So, I demonstrate the influence of the number of samples on playout control by changing the number of samples in this figure. The result of experiments shows that there are no improvements by changing the number of samples. Then, I set the number to be 500 because of less load. high low accuracy adaptive stable dynamics

Performance Evaluation(2/3) Hour Algorithm Target Value[%] Packet Loss Ratio[%] Avg. Playout Delay[ms] 8AM 1 95 5.42 247.66 99 1.08 270.11 99.9 0.10 386.23 2 - 277.88 3 0.12 353.08 4 2.96 258.53 The experiment based on the data measured in 8 a.m. also shows proposed algorithm can keep the packet loss ratio close to the desired loss ratio.

Performance Evaluation(3/3) Hour Algorithm Target Value[%] Packet Loss Ratio[%] Avg. Playout Delay[ms] 2PM 1 95 5.32 242.67 99 1.34 267.29 99.9 0.10 414.48 2 - 1.37 260.62 3 0.26 339.79 4 3.08 252.50 The experiment based on the data measured in 2 p.m. shows the same results. Consequently, we can say that proposed algorithm can control the quality of real-time applications by changing the target value.