Analysis and Evaluation of Skype and Google Talk
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Analysis and Evaluation of Skype and Google Talk Batu Sat and Benjamin W. Wah Department of Electrical and Computer Engineering and the Coordinated Science Laboratory University of Illinois at Urbana-Champaign Ref: B. Sat and B. W. Wah, “Analysis and Evaluation of the Skype and Google Talk VoIP Systems,” Proc. IEEE Int’l Conf. on Multimedia and Expo, July 2006.
Problems Studied z Motivations z Skype (1.3.0.65) & Google Talk (Beta) are widely used but proprietary z Understanding their limitations help design better VoIP systems z Problem statement z Study the various components of Skype and Google Talk z Identify their limitations z Propose new VoIP system z Adapt better to a large range of network conditions July 9-12, 2006 ICME 2006 2
Outline z Experimental Setup z Network evaluations z Architecture of VoIP Systems z Speech Codecs z Packetization z Play-out Scheduling z Loss Concealments z Experimental Results z Future Work July 9-12, 2006 ICME 2006 3
Network Traces from PlanetLab Collecting Internet traces from various geographical locations via PlanetLab nodes Intra-continental paths: • Intra-Asia, intra-Europe, intra-America Inter-continental paths: • North America – Europe/Asia/South America • Asia – Europe/South America July 9-12, 2006 ICME 2006 4
Observations z Highly varying packet delays and jitters ⇒ dynamic jitter control z < 50ms for most intra-continental paths z > 250 ms for some inter-continental and intra-Asian paths z Highly varying packet loss rates ⇒ dynamic loss concealment z < 1% for most intra-continental paths z > 33% for some inter-continental and intra-European paths z Fast changing loss rates ⇒ real-time estimation & frequent feedbacks July 9-12, 2006 ICME 2006 5
Connection with High Loss Rate Bursty Losses: Single: 12.8%, Double: 7.8%, Triple: 0.7%, Quadruple: 12.2% July 9-12, 2006 ICME 2006 6
Connection with Very Low Loss Rate Bursty Losses: Single: 1.5%, Double: 0.1%, Triple: 0.1% July 9-12, 2006 ICME 2006 7
Experimental Setup 4 Troll: Emulating Network Conditions 3 5 VoIP Client (A) VoIP Client (B) 2 Router (C) 6 7 1 Perceptual Speech Quality PESQ 8 July 9-12, 2006 ICME 2006 10
Outline z Experimental Setup z Network evaluations z Architecture of VoIP Systems z Experimental Results z Future Work July 9-12, 2006 ICME 2006 11
Speech Codecs ITU G.729 low bit-rate codec z High efficiency by removing redundancies across frames z Long propagation of errors once a frame is lost Internet Low Bit-Rate Codec (iLBC—RFC3951) z Encode self-decodable frames and remove internal states in receiver z 66 - 90% higher bit-rate than G.729 z Achieve shorter recovery time and robustness to losses z Inadequate without redundancies July 9-12, 2006 ICME 2006 12
Performance Under IID Losses PESQ G729 3-way 24 Kbps iLBC 2-way 26.6 Kbps G729 3-way Sound (OLE2) is better than Wave Sound Wave Sound iLBC 2-way with similar Sound (OLE2) bit rate For loose end-to-end delay conditions, G729 performs better than iLBC for LR>10% July 9-12, 2006 ICME 2006 13
Packet Delays and Redundancy • Skype and Google Talk use 60-ms packet period z Skype packs two 30-ms iLBC frames in a packet (14.4 Kbps) z Google Talk packs variable bit-rate PCM (24 Kbps) Frame Per Packet No 2-way 3-way 4-way Packet Period Redundancy 1 30 ms 0 ms 30 ms 60 ms 90 ms 2 60 ms 30 ms 90 ms 150 ms 210 ms • 3- and 4-way redundancies • Cannot be used with 60-ms periods • Can be used with 30-ms periods without extensive delay Loss rate not affected by reducing packet period from 60 to 30 ms July 9-12, 2006 ICME 2006 14
Play-out Scheduling Skype uses simple FIFO play-out buffer z 1st packet not received 60-ms later than the expected arrival time z Apply iLBC-level loss concealment (RFC3951) z Subsequent packets not received beyond the 60-ms expected time z Played-out when received, causing the whole speech to shift Time 1 3 2 4 1 3 2 4 Original Waveform Degraded Waveform Google Talk employs a jitter buffer similar to that of Skype July 9-12, 2006 ICME 2006 15
Skype: Loss Adaptation & Concealment Feedback Control z Packets in Skype not arriving 600 ms beyond expected time are considered lost z Receiver sends decision to increase or decrease redundancy degree to sender every 2 seconds 5 Bytes 2 sec 2 sec July 9-12, 2006 ICME 2006 16
Skype: Adaptation to Loss-Rate Changes z Loss rate: Step function [0% - 15% - 0%] 5 min steps Faster Increase Slower Decrease Redundancy Degree Very slow loss adaptation due to • Small step sizes to increase/decrease redundancy degrees • Infrequent feedbacks z Loss rate: Step function [0% - 25% - 0%] 5 min steps • Can only track loss-rate changes of 1% (in 2 seconds) 60 sec 200 sec Time July 9-12, 2006 ICME 2006 17
Google Talk: No adaptation to Loss-Rate Changes z No change in packet period and payload size with loss-rate changes z No network feedback 0% 5% 10% 15% 20% 25% 30% 40% 50% July 9-12, 2006 ICME 2006 18
Proposed iLBC Prototype z 30-ms packet period z Packets are sequenced z 1-way to 4-way redundancy decision made at receiver z Relayed with each reverse-path UDP speech packet z Extra 90ms delay for 4-way concealment in the most conservative case z Packets late over 60ms are discarded z Near-term improvement opportunities z Decide on redundancy rate based on loss rate, UCFLR (unconcealable frame loss rate), jitter-buffer size z Relay decision when condition changes (repeated to assure reception) z Adaptively change play-out schedule to minimize delay with a prescribed level of UCFLR July 9-12, 2006 ICME 2006 19
Outline z Experimental Setup z Network evaluations z Architecture of VoIP Systems z Speech Codecs z Packetization z Play-out Scheduling z Loss Concealments z Experimental Results z Future Work July 9-12, 2006 ICME 2006 20
Internet Traces (#2): Low Loss Rate July 9-12, 2006 ICME 2006 21
Internet Traces (#5): High Loss Rate July 9-12, 2006 ICME 2006 22
Conclusions z Internet delays and losses are diverse z Skype and Google Talk z Too conservative loss adaptation and feedback control z Ineffective play-out scheduling z Proposed prototype z More aggressive frame rate and feedbacks z 1-4 ways of redundancy July 9-12, 2006 ICME 2006 23
You can also read