Quality of Service Guarantee over 802.11 Wireless LAN

1. Quality of Service Guarantee over 802.11 Wireless LAN Tzi-cker Chiueh

2. Introduction • Multimedia applications requires QoS support, specifically bandwidth guarantee • But radio link is a shared resource and its access is through CSMA/CA protocol packet collision on the channel • Collision causes two problems: • Non-deterministic access delay • Lower effective link throughput Wireless LAN QoS

3. Mobile Terminal Mobile Terminal 802.11 802.11 Access Point Access Point Wired Ethernet Router Router Media Server Conference Server WLAN Configuration Wireless LAN QoS

4. Quality of Service Metrics • Bandwidth • Delay • Delay jitter • Packet loss rate Wireless LAN QoS

5. Theory • Fluid fair queuing: standard weighted round robin with two exceptions: • Infinitesimally small granularity • Simultaneous service • No queuing delay • Impossible to impelement in practice Wireless LAN QoS

6. Approximation I • Packetized weighted fair queuing (WFQ) • Simulate FFQ by computing the virtual finish time of an incoming packet, and servicing packets based on finish time order VFT(i) = Max{VFT(i-1), VAT(i)} + Packet_Size/BW Delay Bound: Burst/BW + SUM(Packet_Size/BW + Packet_Size/Capacity) Wireless LAN QoS

7. Virtual Time • In each virtual time unit, each backlogged connection i gets a BWi share • Allows the finish time of a packet to be independent of existence of other connections • The number of real time units required to service a virtual time unit of work depends on the number of backlogged connections • VT-RT mapping requires O(N) overhead because of iterative deletion  Various approximations, e.g., SCFQ Wireless LAN QoS

8. Virtual-Real Time Mapping VT-RT mapping overhead could be spread out when queues evolve from backlogged to non-backlogged Virtual Time Real Time Wireless LAN QoS

9. Real Bottleneck • VT-RT mapping overhead is probably not that important in practice • VFT sorting takes O(logN) and is the real scalability limit • Can “locality” help? • How big can N be realistically? Wireless LAN QoS

10. Approximations II • Weighted round robin (WRR) • Simple to implement • Cycle time: tradeoff between efficiency and delay bound • Variations: • Deficit RR • Discrete Fair Queuing: non-packet-based • Smooth RR: still packet-based Wireless LAN QoS

11. Deficit Round Robin • Allow unused credit from previous cycles to carry over Credit = Prev_Credit + BW * Elaspe_Time If (PacketExists and Packet_Size < Credit) Transmit Packet; Credit = Credit – Packet_Size Credit = Cap(Credit) Wireless LAN QoS

12. Discrete Fair Queuing • Discard packet-based assumption • FFQ with small scheduling quantum • Rely on link-layer multiplexing/demultiplexing support • O(1) implementation complexity • Delay bound is proportional to quantum size • Easy to implement in hardware Wireless LAN QoS

13. Smooth Round Robin • Schedule across multiple (M) cycles of WRR • Assume weights are Wi, then M =GCD(Wi) • NxK scheduling matrix, where K = log(M) • Each of M slots is marked with one of the K labels and the distance between consecutive slots marked with the d-th label is 2K-d slots • O(1) complexity and pretty good delay bound compared to WFQ Wireless LAN QoS

14. QoS on WLAN • A wireless channel vs. a wired link • Queues are fundamentally distributed • Raw bandwidth from the AP to different wireless stations may be different • Raw bandwidth from the AP to the same wireless station may be different at different points in time • Interactions with media access control protocol • Hidden node problem Wireless LAN QoS

15. Wireless Rether • Rether is a software-only token passing protocol originally developed for shared-segment Ethernet  adapted to WLAN • Provides bandwidth guarantee to individualapplications, both upstream & downstream • Requires changes to AP and every wireless node • No changes to applications are required • Interoperable with wired network’s DifferServ or 802.11p mechanisms Wireless LAN QoS

16. Wireless Rether • A WLAN node can send traffic only when it receives the token • Token circulates among real-time (RT) nodes in a periodic fashion • Token holding time depends on the total bandwidth reservation on each node • Whatever residual cycle time left by RT nodes are used by the NRT nodes • Requires explicit registration from WRC with WRS Wireless LAN QoS

17. Link Scheduling • DRR but based on channel usage rather than number of bits transmitted • Per-connection packet queuing on each node • Need to dynamically estimate and measure per-packet channel usage time • Overflowed packets are redirected to NRT queue • How many NRT packets should be allowed to be dispatched at a time? Based on global knowledge of NRT queue lengths Wireless LAN QoS

18. Architectural Decisions • Hardware vs. Software implementation • Peer-to-peer vs. Centralized token passing • Essentially the polling mode in 802.11 standard • Is it necessary in infrastructure mode? • Work-conserving vs. Non-work-conserving network link scheduling • To ACK or Not to ACK • May not be necessary always • Implicitvs. Explicitbandwidth reservation Wireless LAN QoS

19. Wireless Rether Client Wireless Rether Client Wireless Rether Client 802.11 Wired Network Access Point Wireless Rether Server Router Rether System Architecture Wireless LAN QoS

20. Bandwidth Reservation • Reservation policy table SrcAddress/Mask, DestAddress/Mask, SrcPortRange, DestPortRange, Bandwidth Requirement • Statistical admission control: based on actual usage rather than reservation sum • Leave slack to avoid starvation of NRT traffic • Automatic two-way reservation for TCP • Intra-LAN connection requires twice the amount of required bandwidth reservation • Special packet queues for Rether packets and other network control packets (ARP and ICMP) Wireless LAN QoS

21. Transparent Packet Scheduling Wireless LAN QoS

22. Wireless Rether Client Wireless LAN QoS

23. Wireless Rether Server Wireless LAN QoS

24. Prototype and Test-bed • Implemented under Red Hat 7.0 • WRS is a 400-MHz Pentium-II machine with 128 Mbytes of memory • WRC is 650-MHz Pentium-III portable machine with 64 Mbytes of memory • Orinoco wireless LAN cards and access point (AP-1000) • Wired network is Fast Ethernet Wireless LAN QoS

25. 2 upstream and 1 downstream Packet size: 1444 bytes Cycle time: 33 ms Wireless LAN QoS

26. Three senders 1.1Mbps sending rate Wireless LAN QoS

27. Throughput vs. Number of Clients Cycle time: 33 ms 1444 bytes 812 bytes 172 bytes 64 bytes Wireless LAN QoS

28. 16Kbps 84Kbps 300Kbps 1Mbps Wireless LAN QoS

29. Improvements • WRS can be readily used as a traffic manager for downstream traffic on a wireless LAN; no WRC is needed on the mobile terminal • TCP-aware good-put management • Automatic content-based bandwidth reservation • Low-latency hand-off for infrastructure-mode wireless LAN, from 2-3 sec to under 100 ms • Vertical hand-off between 802.11b and 2G/GPRS/3G networks • Porting to 802.11a is straightforward • Leveraging 802.11e standard Wireless LAN QoS

30. In Retrospect,…. • Major performance problem lies in token passing overhead due to buffering delay at access points; scheduling and buffering cause no performance problems • Redundancy between link-layer, WRether-layer and network layer mechanisms: registration and ACK • How to leverage MAC-layer header information: • Eliminate token ACK overhead • Turn on the token passing mechanism only when necessary: determine the extent of collision • Trade off between degree of QoS guarantee and QoS mechanism overhead Wireless LAN QoS