Bayesian Piggyback Control for Improving Real-Time Communication Quality Wei-Cheng Xiao 1 and Kuan-Ta Chen Institute of Information Science, Academia Sinica 1 (Now studies in Department of Computer Science, Rice University) Presented by Yu-Chun Chang (National Taiwan University) 2011/5/10CQR 2011 / Yu-Chun Chang1
Outline Motivation Methodology: bayesian piggyback control Performance evaluation Summary 2011/5/10CQR 2011 / Yu-Chun Chang2
Real-Time Multimedia Communication Applications (1/2) Online games Voice chat 2011/5/10CQR 2011 / Yu-Chun Chang3
Video conferences 2011/5/10CQR 2011 / Yu-Chun Chang4 Real-Time Multimedia Communication Applications (2/2)
Motivation Popular real-time and interactive applications – Real-time network games, VoIP Quality of Experience 2011/5/10CQR 2011 / Yu-Chun Chang5
Traffic Pattern of Real-Time Communication Real-time – Small packet size – High packet rate User friendship – Low end-to-end delay 2011/5/10CQR 2011 / Yu-Chun Chang6
User Satisfaction Key Factor Smoothness of data communication Long end-to-end delay – Packet loss – Retransmission Jitter – Network congestion – Packet reordering 2011/5/10CQR 2011 / Yu-Chun Chang7
Real-Time Communication Mechanism 2011/5/10CQR 2011 / Yu-Chun Chang8 Detect loss events Retransmit loss packets
A GOOD Real-Time Communication Mechanism Should … 2011/5/10CQR 2011 / Yu-Chun Chang9 Work without modifying inherent network protocol and designs Decide whether a packet has been lost before the retransmission timer expires Avoid generating too much unnecessary traffic
Contributions We design a packet loss event detector to – detect packet loss events without modifying protocol – determine packet loss events before retransmission timer expires – avoid unnecessary transmission overhead 2011/5/10CQR 2011 / Yu-Chun Chang10
Outline Motivation Methodology: Bayesian piggyback control Performance evaluation Summary 2011/5/10CQR 2011 / Yu-Chun Chang11
Methodology Bayesian Piggyback Control – Bayesian inference – Probability density function estimation – Piggyback scheme 2011/5/10CQR 2011 / Yu-Chun Chang12 Detect loss event Retransmission mechanism
Concept 2011/5/10CQR 2011 / Yu-Chun Chang13 Packets are transmitted via intermediate routers Drop-tail queue
Packet Round-Trip Time 2011/5/10CQR 2011 / Yu-Chun Chang14 Summed up from – processing delay – propagation delay – queueing delay RTT is mainly related to this Drop-tail queue
Drop-Tail Queue 2011/5/10CQR 2011 / Yu-Chun Chang15 TailHead Enqueue FULL Drop packet! Relay packet Suffers the longest queueing delay
Bayesian Inference (1/2) 2011/5/10CQR 2011 / Yu-Chun Chang16
Bayesian Inference (2/2) 2011/5/10CQR 2011 / Yu-Chun Chang17
Probability Density Function Estimation The Histogram-based method – A simple and intuitive method to estimate the conditional probability mass function The Parzen method – A more sophisticated method to smooth the curve of the histogram-based method 2011/5/10CQR 2011 / Yu-Chun Chang18
The Histogram-Based Method 2011/5/10CQR 2011 / Yu-Chun Chang19
The Parzon Method 2011/5/10CQR 2011 / Yu-Chun Chang20
Detection Tradeoff False positive rate (FPR) – An event successful is judged as lost – FPR ↑ : some additional traffic will be injected False negative rate (FNR) – An event lost is judged as successful – FNR ↑ : will cause very high delay 2011/5/10CQR 2011 / Yu-Chun Chang21
Penalty Strategy 2011/5/10CQR 2011 / Yu-Chun Chang22
Piggyback Scheme Previous data considered lost will be appended to construct a new packet Retransmit loss data before transport layer timer expires Advantages – Reduce the bandwidth requirement for packet headers – Decrease network overheads 2011/5/10CQR 2011 / Yu-Chun Chang23
Bayesian Piggyback Control 2011/5/10CQR 2011 / Yu-Chun Chang24
Outline Motivation Methodology: Bayesian piggyback control Performance evaluation Summary 2011/5/10CQR 2011 / Yu-Chun Chang25
Simulation Setup (1/2) Simulator: ns2 Network topology: transit-stub graph 50 nodes: 1transit domain / 6 stub domains Communication server: 1 node Hosts running real-time applications: 15 nodes Hosts generating cross traffic: 34 nodes 2011/5/10CQR 2011 / Yu-Chun Chang26
Simulation Setup (2/2) Average bandwidth – Transit-transit domain: 2000 KB/sec – Transit-stub domain: 2000 KB/sec – Stub-stub domain: 1000 KB/sec Real-time communication applications – Packet rate: 30 ms a packet – Packet size: 100 ~ 300 bytes Cross traffic: UDP packets (750 KB/sec per host) 2011/5/10CQR 2011 / Yu-Chun Chang27
Detection Accuracy 2011/5/10CQR 2011 / Yu-Chun Chang28
Performance Metric ROC (Receiver Operation Characteristics) – TPR: true positive rate – FPR: false positive rate 2011/5/10CQR 2011 / Yu-Chun Chang29
ROC Curve 2011/5/10CQR 2011 / Yu-Chun Chang30 20%
Effect of Piggyback Scheme Two kinds of delay affect user satisfaction – End-to-end delay – Lag (time difference of two contiguous message) Performance comparison – Optimistic mechanism Original system without any loss detection or retransmission mechanism – Pessimistic mechanism A message will always be retransmitted until it is received 2011/5/10CQR 2011 / Yu-Chun Chang31
End-to-End Delay Analysis CQR 2011 / Yu-Chun Chang Most e2e delay values are below 1 sec. The probability is higher than 99%
Lag Analysis CQR 2011 / Yu-Chun Chang Our method achieve lags about only 25%-50% of those of the optimistic mechanism.
Overhead Observation
Summary Demand for real-time communication applications significantly increases Long delays degrade users’ satisfaction Packet loss events trigger time consuming timeout retransmission mechanism We proposed Bayesian Piggyback Control to judge packet loss events before the retransmission timer expires and retransmit loss packets efficiently (with few overheads) The proposed detector achieves at least 80% detection rate as the false alarm below 20% 2011/5/10CQR 2011 / Yu-Chun Chang35
Thank you for your attention! 2011/5/10CQR 2011 / Yu-Chun Chang36