1 Andreas Koepsel, Adam Wolisz Voice transmission in an IEEE 802.11 WLAN based access network Andreas Koepsel #, Adam Wolisz Telecommunication Networks Group Technical University Berlin www-tkn.ee.tu-berlin.de # Now co-founder of Acticom R&D (Start-up fever in Berlin!) [email protected]

2 Andreas Koepsel, Adam Wolisz Motivation Several wireless technologies are competing for wireless access in local environments –DECT, GSM –Hiperlan/2 –Bluetooth –IEEE 802.11 Usage of a single technology for voice and data avoids expensive parallel infrastructures Suitability of IEEE802.11 WLAN for parallel use for data and voice streams Distributed vs. Centralized coordination schemes

3 Andreas Koepsel, Adam Wolisz Architectural Scenario IEEE 802.11 WLAN cells used as last hop within wired local area networks, both for Data and voice All mobile nodes (MN) create best-effort traffic A subset of the wireless terminals generates voice calls to destinations outside of the local environment

4 Andreas Koepsel, Adam Wolisz Medium Access in IEEE 802.11 (DCF) fixed

5 Andreas Koepsel, Adam Wolisz Medium Access in IEEE 802.11 (PCF) fixed

6 Andreas Koepsel, Adam Wolisz Audio source modeling and QoS requirements G.711 (PCM) /G 723.1 codecs Integrated silence detection => 2-state Markovian chain simulating ON-OFF-sequences with exponential distributed periods QoS requirements –maximum end-2-end delay limitations of ~250-300ms –Maximum loss rate rate (dependent on coding). Queueing and Dropping: –An audio packet that exceeds its maximum lifetime is dropped hence reducing the overall offered load –The lifetime varies for different audio connections (different destinations)

7 Andreas Koepsel, Adam Wolisz Best Effort / Physical Layer Best effort traffic –Packet interarrival times from Pareto distribution –Packet length distribution from packet trace (Harvard university) –Packet length maximum 1500bytes Transmission systems –IEEE 802.11b DSSS system with up to 11Mbit/s –IEEE 802.11a OFDM system with 24Mbit/s mandatory and up to 54Mbit/s optional Gilbert-Elliott- based channel error model independent channels among every pair of mobiles

8 Andreas Koepsel, Adam Wolisz Audio transmission with DCF – Delay(1) A minimum transmission link speed of 11Mbit/s is required, further increase of link speed very advantegous... Data Traffic – 12 Nodes, Voice Traffic:4VoIP flows,  6Mbit/s Link speeds:2, 11, 24 and 54Mbit/s Bit error rates from 10 -7 up to 10 -3

9 Andreas Koepsel, Adam Wolisz Increasing the link speed- just the solution?  Fixed length PHY PDU header sent with a reduced basic rate (1Mbit/s for DSSS, 6Mbit/s for OFDM), fixed interframe spaces  Link speed enhanced by factor 5 11Mbps  54Mbps  Improvement seen by application (  2,2) 45%| 11Mbps  20%| 54Mbps

10 Andreas Koepsel, Adam Wolisz Audio transmission with DCF (3)  R-time traffic is mixed up with best-effort packets, why not separate?  The mean channel access delay decreases significantly without the need of a high complexity!  Mean channel access delay drops at 11Mbps from 90ms to 50ms at BER 10 -4 Single queue Two queues

11 Andreas Koepsel, Adam Wolisz Audio transmission with PCF  The superframe interval is chosen according to the used audio codec ( eg. with G.723.1, 5.3kbps, frame interval 30ms)  PCF offers significantly lower channel access times ... and an improved probability density function (see next slide) ... but reduces the overall throughput for best-effort traffic

12 Andreas Koepsel, Adam Wolisz Probability density function of Channel Access Delay at 11Mbps PCF Probability density function of Channel Access Delay at 11Mbps - DCF  

13 Andreas Koepsel, Adam Wolisz Additional Improvement Due to existence of talkspurts the application will have brakes in sending...lost cycles. An idea: Why not remove the station from the polling cycle if there are no data, and introduce it again after a first packet has been tranmitted in DCF? But: there might be losses of the polling frame... RTP fileds SSRF id (for flow identification) and the MORE flag migh be used to decide about the continuation of the flow.

14 Andreas Koepsel, Adam Wolisz  Avoiding unneccesary polling attempts caused by talkspurts  Implicit signalling by a simple MORE flag when further audio packets will be available 

15 Andreas Koepsel, Adam Wolisz Conclusion A simple separation of the transmission queues within DCF decreases the channel access times significantly. DCF wastes capacity when used with improved coding schemes (higher bit rates!). PCF provides lower delay and lower delay variance to audio flows as compared to DCF. The PCF overhead plays a less significant role for higher data rates. PCF performance might be further improved by introducing a flow aware scheduling policy.

16 Andreas Koepsel, Adam Wolisz Further Research... Introduction of true elastic load for best effort traffic. More insight in different load scenarios (best effort vs. RT, number of active stations) Differentiation of packet importance for improved scheduling... Papers in ACM MM2001 ans SPIE Denver, August 2001

17 Andreas Koepsel, Adam Wolisz Example: Voice source  Time aspect - G.711  PCM 64kbps source - G.723.1  5.3 and 6.4kbps speech codecs - audio flow (in time) can be fairly well simulated by a two state Markov chain with geometric distributions (talk- spurts) - states T ALK and S ILENCE with means TALK =1.35 ms and SILENCE =1.15 ms  Importance of packets... - not identical! - position of the lost packet influences essentially the quality -„Voiced“ frames: important, decoder needs more time to synchronize after loss.... -Loss of a „voiced“ frame at the border „unvoiced“-“voiced“ even worse. -How does it work for wired Internet

18 Andreas Koepsel, Adam Wolisz QoS – Interactive Audio Classical wisdom  Audio connections need a strictly limited delay/delay variation. Overdue packets contribute to loss rate...  Audio quality decreases significantly if the number of packet losses exceeds some level (n %) Priority in preserving packets  SPB: Speech property based prioritization  ALT: just alternating high/low prioritization  FEC: additional FEC protection

19 Andreas Koepsel, Adam Wolisz Speech quality - wired internet excellent good fair 0 0.05 0.25 SPB-FEC SPB-DIFFMARK ALT-DIFFMARK NO MARK FULL MARK drop probability p 0 EMBSD Perceptual Distortion 5