The Effect of First-Hop Wireless Bandwidth Allocation on End-to-End Network Performance

1. The Effect of First-Hop Wireless Bandwidth Allocation on End-to-End Network Performance Lili Qiu, Paramvir Bahl, Atul Adya Microsoft Research NOSSDAV’2002 Miami Beach, Florida

2. Outline • Motivate the problem • First hop wireless bandwidth reservation schemes • Performance evaluation • Related work • Conclusion

3. Introduction • IP telephony is becoming popular • With the growth of wireless technologies, an IP-based portable phone is a compelling device for voice communication

4. AudioVox Thera (CDMA2000) - PPC 2002 OS Siemens SX 45 (GPRS) - PPC 2002 OS Nokia 9110 (GSM) - GEOS OS Kyocera QCP 6035 (CDMA)- Palm 3.5 OS MSR’s UCoM (802.11) - PPC 2002 Handspring Treo 180 - Palm OS (GSM) MS SmartPhone - WinCE 3.0 OS Samsung I300 - Palm OS 49 million PDA-Phones by the year 2007 [Cellular News 1/23/02]

5. Research Issues • Design space • Power • Security • Mobility • Bandwidth  this paper’s focus • … • Wireless bandwidth management • Wireless technologies used in the first hop are still slow • Challenges • End-to-end reservation is preferred for ensuring QoS • Little QoS support in the Internet  reservation only at wireless hops • Effectiveness of local reservation depends on location of bottleneck

6. Motivation:Throughput of Internet Paths • How often is a wireless application traversing the Internet rated-limited by its Internet path rather than wireless hop? • Analyze the Internet tcpdump traces collected at microsoft.com • Incoming & outgoing web traffic, software download traffic, streaming media traffic

7. Throughput of Internet Paths • Different clients experience widely different throughput, from 1 Kbps to 10 Mbps • Over 30% clients have throughput less than 20 Kbps Internet path can become bottleneck • Useful to consider congestion level of the Internet when making the bandwidth allocation decision in the first hop

8. Temporal Stability of Internet Throughput Over 90% of the hosts have throughput variation within a factor of 2  Internet throughput is stable

9. Our Approach • Observation • Not efficient to reserve more bandwidth than what an application would use. • An application may use less BW • either because it generates data at slower rate • or because bottleneck is at other links (e.g. Internet). • Approach • Passively monitor applications’ bandwidth • Adaptively modify allocated wireless bandwidth according to usage • Places to deploy the technique • Infrastructure (Access point/Access server) • Client (PDA)

10. First-hop Wireless Bandwidth Reservation Schemes • No reservation: best effort • R0: Reserve s • R1: Reserve min(s, f*I) • R2: Same as R1, except it periodically re-adjusts allocation s: the rate specified by the source I: Internet bandwidth f: tolerance factor to account for estimation error

11. Performance Evaluation Receivers • Simulations in ns-2 • Senders use TFRC [FHP00] • Compare the above four reservation schemes • Three scenarios • Congestion at wireless hop • Congestion at the Internet path • Congestion at both places Senders 6Mbps A B wireless hop Internet path

12. Scenario 1: Congestion in the wireless hop • Simulation scenario: • S = 48 Kbps • I = 96 Kbps • s: the rate specified by the source • I: Internet bandwidth • Congestion in the wireless hop First hop reservation maintains QoS when congestionoccurs in the first hop.

13. Scenario 2: Congestion in the Internet path Simulation scenario: S = 48 Kbps, I = 24 Kbps  congestion in the Internet • R0 is wasteful. • First hop reservation is ineffective when congestion occurs in the Internet.

14. Scenario 3: Congestion at both Internet and wireless hop S = 48 Kbps, I = 24 or 96 Kbps  congestion at both places Reservation based on Internet throughput performs the best.

15. Scenario 3: Congestion at both Internet and wireless hop (Cont.) • Use the throughput in the Internet trace for our simulation • Pick hosts from the Dec. 2000 throughput trace, and assign their perceived throughput to the bandwidth of the Internet path • Vary the bandwidth of the links according to the trace • Estimate throughput using past measurements • Tolerance factor (f) = 1.5 • Desired sending rate of a source • Either CBR: 16 Kbps, 24 Kbps, 32 Kbps, 48 Kbps • Or VBR: the rate of video traces we collected • Poisson arrival & departure • mean duration = 8 minutes

16. Scenario 3: Congestion at both Internet and wireless hop (Cont.) Allocating bandwidth that adapts to the Internet path’s throughput is even better.

17. System Components Allocation Server • monitors throughput and re-adjusts reservation periodically Access Server • Polices users

18. Related Work • Studies on Internet path properties • Internet throughput remains stable on the time scales of minutes [BSSK97, ZDPS01] • Admission controls • RSVP [ZDES+93] • Measurement-based admission control • Endpoint admission control [BKSS+00] • Wireless QoS • IEEE 802.11e • Subnet bandwidth manager [RFC 2814]

19. Conclusions • Provide applications with better QoS without infrastructure support in the Internet • Study several bandwidth allocation techniques for wireless hops • Adaptive bandwidth allocation for first-hop wireless based on passive observation of Internet paths performs the best • Has better quality than no reservation • Admits more flows than naïve reservation • Design choices • Applications • Wireless real time applications

20. Acknowledgement • Venkata N. Padmanabhan • Scott Hogan • Rob Emanuel • Chris Darling • Al Lee