Purdue University Computer Science. End-to-End Transport of Real-Time Trac. Using. Kihong Park. Purdue University. West Lafayette, IN 47907

Transcription

1 End-to-End Transport of Real-Time Trac Using Adaptive Forward Error Correction Kihong Park Department of Computer Sciences Purdue University West Lafayette, IN 4797

2 Overall Goal Facilitate transport of real-time trac e.g., video, audio, voice, interactive applications such that QoS-sensitive end-to-end ecient adaptive

3 Problem Hard real-time application: Sender Receiver a b c d e f g a b d e Network t f g t packet drop queueing delay

4 Thus RTT can exceed time constraint retransmission (ARQ) infeasible Solutions: resource reservation & admission control { overprovisioning { inecient due to self-similar burstiness forward error correction (FEC),! proactive

5 Forward Error Correction 2 k 2 k E k+ k+h k k + h Network 2 k 2 k D k+ k+h k > k - k data packets encoded as n = k + h code packets transmit n code packets receipt of any k packets allows for recovery

6 Features: packet-level FEC E, D with \k-out-of-n" property exist,! e.g., Reed-Solomon, IDA Dierence with traditional FEC: packet-level vs. bit-level independent vs. correlated information loss queueing-induced correlation self-similar burstiness

7 Eective QoS-control using FEC requires: tolerance to burstiness adaptability to network state { if network is \good," inject low redundancy { if network is \bad," inject high redundancy,! adaptive FEC,! maintain invariant target QoS Caveat: Injecting \too much" redundancy can be counterproductive,! i.e., eventually can decrease QoS

8 Unimodal redundancy-recovery relation: redundancy h sender transmits block ofn= k+ hpackets ( n) packets arrive timely at receiver γ γ * γ * h * h * h

9 Features: is related to perceived QoS maximum recovery target recovery Optimal control problem: given (user specied) adjust h such that = Control law: dh dt = f(desired QoS; network state)

10 Stability Analysis Control laws of the form: if <, increase h if >, decrease h γ γ = γ* *. h * = h * h

11 γ γ * γ *. h * h * h γ γ * γ *. h * = h * h

12 AFEC Protocol Core protocol: ( dh dt = (, ); if d=dh,,ah; otherwise. Main feature: inside stable regime: symmetric control inside unstable regime: exponential backo

13 Real-Time MPEG Video Transport teleconferencing application AFEC customization to MPEG video payload end-to-end QoS-sensitive transport implemented entirely in software Sender Receiver I, P, B,... FEC encoder Network FEC decoder C s h γ γ Cr I, P, B,... MPEG II player

14 Forward error correction using IDA Encoding: l s = l / k... 2 k M A B... = X (k X s) (k + h X k) (k + h X s)

15 Decoding: encoding X = AB decoding A, k X k = B

16 Special features of MAFEC: receiver-oriented QoS control Stringency control dierentiated weighting

17 Experimental Set-Up AFEC sender router AFEC receiver cross traffic MAFEC application (sender/receiver) cross trac source congurable router UltraSparc workstations FastEthernet (Mbps)

18 Sample MPEG-I video traces: traffic volume (bytes) Simpsons MPEG video time (msec) 8 terminator traffic volume (bits) time (min)

19 Performance Measurements Redundancy and QoS: hit rate.95 hit rate (Simpsons MPEG) redundancy h.6.5 packet loss rate (Simpsons MPEG) packet loss rate redundancy h

20 hit rate I frame P frame B frame redundancy h

21 Static FEC vs. AFEC: packet drop hit packet drop 2.5 hit frame number frame number packet drop hit packet drop 2.5 hit frame number frame number

22 AFEC dynamics:.95 measured hit rate target hit rate hit rate frame hit 2.8 hit frame 2 redundancy h 8 h time

23 .95 measured hit rate target hit rate hit rate frame hit 2.8 hit frame 2 redundancy h 8 h frame

24 Backo instances: backoff 2.8 backoff frame backoff 2.8 backoff frame

25 Operating point dynamics: hit rate static FEC AFEC: target hit rate AFEC: achieved hit rate h hit rate static FEC AFEC: target hit rate AFEC: achieved hit rate h

26 Impact of buer capacity:.2 small buffer (5) medium buffer (55) large buffer (65) hit rate redundancy h,! packet loss vs. delay domination

27 Impact of self-similar burstiness:.9 highly bursty (H =.75) medium bursty (H =.65) weakly bursty (H =.58) hit rate redundancy h.9 highly bursty (H =.75) medium bursty (H =.65) weakly bursty (H =.58) hit rate redundancy h

28 hit rate, packet loss rate hit rate packet loss rate buffer capacity hit rate, packet loss rate hit rate packet loss rate bandwidth

29 Conclusion QoS-sensitive transport of real-time trac using FEC adaptive FEC end-to-end optimal control problem implementation for real-time MPEG video transport software implementation desirable performance characteristics,! real-time MPEG audio for Internet telephony