TCP/IP performance over 3G wireless links with rate and delay variation

1. TCP/IP performance over 3G wireless links with rate and delay variation Mun Choon Chan and Ramjee, Bell Labes Presented by: Jinqiang Yang

2. Outline • Introduction • Channel state based scheduling and variable rate and delay • Performance evaluation • “Ack” regulator to improve TCP/IP performance • Simulation results

3. Introduction • 3G technologies provide wide band wireless data channel. • 3G1X 144Kbps • 3G1X-EVDO / HDR 2Mbps • Wireless channels are unstable / of poor quality • Neighbor channel interference, • Rayleigh fading • Multi-path fading

4. Wireless channel quality • Packet losses are frequent • The TCP perceives packets lost over wireless channel as congestion • Loss over wireless channel decreases TCP throughput • Link layer retransmission / channels state based scheduling are employed to address this problem

5. Link layer retransmission • Link layer retransmission: • Snoop agent: retransmits lost packet on wireless channel, suppresses duplicate Acks • Split TCP connection: the wireless part will do fast recovery and retransmission • Link layer retransmission is now part of the standards of CDMA2000 and UMTS

6. Channel state based scheduling: variable rate and delay

7. Channel state based scheduling • Intelligent packet scheduling mechanism • At the base station, wireless channel states are taken into account when scheduling data packets for different users • Give priority to users with better channel quality • Strict priority can lead to starvation of users with poor channel quality • Long term fairness can be achieved by proportional fair algorithm

8. Channel state based scheduling • The gain It improves overall throughput • The cost Increased delay and rate variability

9. Channel state based scheduling • Rate and delay variability effects: • The “Ack” arrivals are bursty => ( Ack compression) • Bursty Acks leads to bursty packet release from the source => • Then the bust of packets may experience multiple packet losses => • TCP throughput is degraded

10. TCP/IP over wireless channel Performance evaluation

11. Network architecture

12. Experiment • 1000 ping packets were sent over CDMA 1X link. 144Kbps down link, 8 Kbps up link • Delay and Ack arrival time are measured. • No buffer overflow loss, because in this experiment buffer size is larger than the TCP window size.

13. Ping Latency CDF of latency Latency: span from 150ms to 1s

14. Ack inter arrival time CDF of Ack inter arrival time

15. Ack inter arrival time • Evenly spaced Ack should be 172ms apart • 10% of the Ack arrive in 50ms after the previous one • significant Ack compression observed.

16. Effect on Window size

17. Effect on Window size

18. Packet loss and window size

19. Ack regulator

20. Ack regulator • Goal: to achieve saw-tooth congestion window behavior even in presence of varying delay and rate • Method: by controlling the buffer overflow process in bottle-neck link • Ack regulator regulates the Ack flow back to the TCP source

21. Ack regulator • RNC needs to maintain per-TCP flow queues instead of per-user queues. • Ack regulator is designed to avoid any buffer overflow loss until the TCP congestion window size reaches a preset threshold • Over the threshold, allow only a single buffer overflow loss.

22. Ack regulator implementation

23. Ack regulator • Conservative Mode: • Each time an Ack is sent back to the source, there is buffer space for at least two data packets from the source • Ensures that there is no buffer overflow loss even the TCP source increases window size

24. Ack regulator • Non- conservative mode • Number of Acks sent back to source is less than or equal to the maximum packets the free buffer space can hold • A single packet loss occurs if TCP increases its window size by 1

25. Ack regulator process

26. Simulation Results

27. Simulation topology

28. Simulation parameters Single TCP flow n=1 FR=200Kb/s, RR=64Kb/s, FD has exponential distribution with mean between 20ms to 100ms RD= 400ms – mean(FD) Buffer size = 10 packets

29. Variable delay

30. Buffer size

31. Variable delay • When delay variance increases from 20 to 100: • Throughput of Reno decrease by 30% • Throughput of Sack decreases by 19% • Throughput of Reno and Sack with Ack regulator decreases by 8% • Ack regulator is able to maintain a throughput of over 80% of maximum throughput (200Kbps)

32. Variable Bandwidth Simulation parameters: • FR uniformly distributed with mean 200Kbps and variance from 20 to 75 • FD=200ms RR=64Kbps, RD=200ms • Buffer size=10

33. Variable bandwidth

34. Buffer size

35. Variable bandwidth • Compared to TCP Reno, Ack regulator improves throughput by up to 15% • TCP Sack performs very well and has almost the same throughput as Ack regulator • If the rate variance is not large, Sack can handle the variance. For large rate variations, Sack performs better with Ack regulator.

36. Summary • Ack regulator is able to increase TCP Reno and TCP Sack performance • Ack regulator delivers the same high throughput irrespective whether the TCP source is Reno or Sack • Ack regulator showed robust high throughput across different buffer sizes

37. Thanks