1. WHITE – Achieving Fair Bandwidth Allocation with Priority Dropping Based on Round Trip Time Name : Choong-Soo Lee Advisors : Mark Claypool, Robert Kinicki Reader : Craig Wills Date: March 25, 2002

2. Outline Introduction Related Work Approach Evaluation Conclusion

3. Introduction Current internet uses routers with droptail queue management Droptail introduces the problem of global synchronization There are many active queue managements proposed but most of them are concerned with overall throughput and delay but not with fairness Flows are not homogeneous but heterogeneous Robust flows vs. Fragile flows

4. Related Work Random Early Detection (RED) Flow RED (FRED) Core-Stateless Fair Queuing (CSFQ) Deficit Round-Robin (DRR)

5. RED [FJ93] Based on average queue size min th max th queue size 0 1 max_p MinthMaxth

6. FRED [LM97] Modification to RED Maintains per-flow state information

7. CSFQ [SSZ98] Rate-based Active Queue Management Distinguishes between edge and core routers Edge routers label packets Core routers use these labels to treat packets fairly Estimates fair share and uses it to drop packets

8. DRR Implementation of Fair Queuing Maintains per-flow state information

9. Overview Goals Achieve fair allocation close to Fair Queuing and comparable or better than RED, FRED and CSFQ under most scenarios. Reduce complexity by not having to maintain per flow state Per Packet No Per Packet Per FlowNo Per Flow DRRFREDWHITE CSFQ RED

10. Outline Introduction Approach Round Trip Time at the Edge Average Round Trip Time at the Router Drop Probability Based on Round Trip Times Evaluation Conclusion

11. Approach Modification to RED Adjusts max_p per packet Supports both dropping and marking of packets Dropping vs. Marking Dropping WHITE : Chardonnay Marking WHITE : Chablis Round Trip Time at the Edge Average Round Trip Time at the Router Drop Probability Based on Round Trip Times

12. Round Trip Time at the Edge Edge Hint Packets get labeled with additional information We want the lowest RTT as our hint Modification to TCP-Reno with TCP-Vegas RTT Computation 4-17 bits in the IP header available for additional information if no fragmentation [SZ99]

13. Average Round Trip Time at the Router Now that we have the RTT edge hint, RTTs are exponentially weighted (R average ) at the router Due to high fluctuation of R average, we use extra steps to compute stabilized value of RTT (R formula ) How long it has been out of  12.5ms

14. Drop Probability Based on Round Trip Time Now, we want to use RTT edge hint and average RTT at the router to compute drop probability TCP-Friendly Formula [PFK98] Simplify T 1 = T 2

15. Drop Probability Based on Round Trip Time

17. For Chardonnay, 0.71 corresponds to   robust) and 1.58 to  fragile). For Chablis, 1.58 corresponds to both   robust) and  fragile). However, simulation results showed that values of (0.65, 1.4) worked the best for Chardonnay and (1.6, 1.4) for Chablis.

18. min th max th queue size 0 1 max_p WHITE Algorithm q ave robust flow fragile flow

19. Outline Introduction Approach Evaluation Setup Experiments Chardonnay vs. Chablis Conclusion

20. Setup Network Simulator 2 (NS-2) was used to run all the simulations. Modification to source code to include RTT edge hints and to implement WHITE. We ran 6 experiments with RED, FRED, CSFQ, DRR, Chardonnay and Chablis

21. Setup N0N0 N1N1 N2N2 N 29 RD Queue Size: 120 10 Mbps, 5ms 5 Mbps RED/FRED min th :10 max th :30 w q :0.0008 max_p:0.1 WHITE (Chardonnay, Chablis) min th :10 max th :30 W q :0.0008 max_p:0.1  :0.65, 1.6  :1.4, 1.4 CSFQ K:100ms K  :100ms K c :100ms

22. Experiments Uniformly Distributed Latencies (Exp1) Round trip latencies from sources were 20ms, 30ms, 40ms, …, 310ms. Balanced Clustered Latencies (Exp2) Unbalanced Latencies (Exp3, Exp4) Dynamic Latencies (Exp5, Exp6)

23. Uniformly Distributed Latencies

29. Experiments Uniformly Distributed Latencies (Exp1) Balanced Clustered Latencies (Exp2) Unbalanced Latencies 1 flow with 20ms round trip latency and 29 flows with 200ms round trip latency (Exp3) 1 flow with 200ms round trip latency and 29 flows with 20ms round trip latency (Exp4) Dynamic Latencies (Exp5, Exp6)

30. Unbalanced Latencies: 1 Robust vs. 29 Fragile

36. Unbalanced Latencies: 1 Fragile vs. 29 Robust

42. Experiments Uniformly Distributed Latencies (Exp1) Balanced Clustered Latencies (Exp2) Unbalanced Latencies (Exp3, Exp4) Dynamic Latencies 10 flows with 50ms round trip latency, 10 flows with 100ms round trip latency and 10 flows with 200ms round trip latency (Exp6)

43. Dynamic Latencies Robust Average Fragile 0s60s90s120s30s ABCD

44. Dynamic Latencies

50. Overall Comparison

51. Chardonnay (Dropping) vs. Chablis (Marking)

52. ExperimentChardonnayChablis Drop (%)Goodput (Mbps) Drop (%)Goodput (Mbps) 11.809.590.0009.65 22.709.910.0009.98 31.469.670.0009.78 43.599.960.0079.96 52.569.760.0029.85 62.499.690.0039.82

53. Outline Introduction Approach Evaluation Conclusion Future Work

54. Conclusion Performance of Chardonnay and Chablis is better than RED, FRED and CSFQ and comparable to DRR RTT edge hints can be used to approximate DRR’s performance without the complexity of maintaining per-flow state information Marking performed better Less drops Better goodput

55. Future Work Current version of WHITE does not support any non-responsive flows such as UDP flows Adaptive mechanism is necessary to support much more flows than those in simulations