1. Presentation Date : 7 Sep 2005 1 Evaluation and Characterization of Available Bandwidth Probing Techniques Ningning Hu, Department of Computer Science, Carnegie Mellon University, Pittsburgh Peter Steenkiste, Department of Electrical and Computer Engineering, Carnegie Mellon University, Pittsburgh IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 21, NO. 6, AUGUST 2003

2. Presentation Date : 7 Sep 20052 Motivation The questions : What are good techniques for estimating available bandwidth and what factors affect the measurement accuracy are still open questions. Use modeling, measurements and simulations to better characterize the interaction between probing packets and the competing network traffic.

3. Presentation Date : 7 Sep 20053 Contributions Develop a single-hop gap model. The packet pair gap can be used to accurately characterize the competing traffic. Develop two packet pair techniques—initial gap increasing (IGI) and packet transmission rate (PTR)—to characterize the available bandwidth on a network path. Explore how packet train parameters can affect the measurement accuracy. Use simulations to quantify how various factors impact the accuracy of the algorithms in multihop networks.

4. Presentation Date : 7 Sep 20054 Single-Hop Gap Model A.Single-Hop Gap Model

5. Presentation Date : 7 Sep 20055 Single-Hop Gap Model

6. Presentation Date : 7 Sep 20056 Single-Hop Gap Model ---DQR ---JQR

7. Presentation Date : 7 Sep 20057 Single-Hop Gap Model ---IGI formula (for CT) ---PTR formula (for a_bw) The key to an accurate available bandwidth measurement algorithm is to find a g I value so that the probing packet train operates in the JQR. B.Probing Packet Trains

8. Presentation Date : 7 Sep 20058 Single-Hop Gap Model C.Testbed Illustration

9. Presentation Date : 7 Sep 20059 Single-Hop Gap Model 1)Effect of JQR: Capturing Competing Traffic:

10. Presentation Date : 7 Sep 200510 Single-Hop Gap Model

11. Presentation Date : 7 Sep 200511 Single-Hop Gap Model 2)Effect of DQR: Losing Competing Traffic:

12. Presentation Date : 7 Sep 200512 Single-Hop Gap Model Discussion There is a proportional relationship between the output gap and the competing traffic.

13. Presentation Date : 7 Sep 200513 IGI and PTR Algorithms A.Impact of Input Gap g I

14. Presentation Date : 7 Sep 200514 IGI and PTR Algorithms B.IGI and PTR Algorithms In their experiments, δis set to 0.1

15. Presentation Date : 7 Sep 200515 Comparative Evaluation A.Network Paths Use the following 3 methods to measure the available bandwidth: 1.IGI and PTR 2.Pathload 3.Bulk data transfer (Iperf)

16. Presentation Date : 7 Sep 200516 Comparative Evaluation Note: Link capacities are measured by bprob and the RTTs by ping.

17. Presentation Date : 7 Sep 200517 Comparative Evaluation B.Measurement Accuracy

18. Presentation Date : 7 Sep 200518 Comparative Evaluation (P1) CORNELL -> MA

19. Presentation Date : 7 Sep 200519 Comparative Evaluation C.Convergence Times

20. Presentation Date : 7 Sep 200520 IGI and PTR Algorithm Properties A.Probing Packet Size

21. Presentation Date : 7 Sep 200521 IGI and PTR Algorithm Properties

22. Presentation Date : 7 Sep 200522 IGI and PTR Algorithm Properties Small packet sizes can result in high errors in the available bandwidth estimation. The resulting probing train is more sensitive to the burstiness of the competing traffic. The small gap values are harder to measure accurately. Probing flows with larger packets are more aggressive than probing flows with smaller packets, so they “observe” a higher available bandwidth. The packet sizes that result in the closest estimates are 500 and 700 Byte.

23. Presentation Date : 7 Sep 200523 IGI and PTR Algorithm Properties B.Packet Train Length and Number of Probing Phases NWU -> CMU 100Mbps CORNELL -> CMU 10Mbps

24. Presentation Date : 7 Sep 200524 IGI and PTR Algorithm Properties NWU -> CMU 100Mbps CORNELL -> CMU 10Mbps

25. Presentation Date : 7 Sep 200525 Multihop Effects

26. Presentation Date : 7 Sep 200526 Multihop Effects A.Tight Link Is Not the Bottleneck Link

27. Presentation Date : 7 Sep 200527 Multihop Effects B.Interference From Traffic on “Nontight” Links Competing traffic before the tight link has almost no impact on the accuracy. Changes in gap values before the tight link can be reshaped by the router. Competing traffic after the tight link does have an effect and, its impact increases with the level of competing traffic.

28. Presentation Date : 7 Sep 200528 Multihop Effects C.Timing Errors Measurement errors on the destination side will have a more significant impact since they will directly change the gap values that are used in the IGI and PTR formulas.

29. Presentation Date : 7 Sep 200529 Conclusion The authors present a simple gap model to identify under what conditions packet pairs probing yields useful information about the available bandwidth along a path. They design IGI and PTR for estimating available bandwidth based on the gap model. IGI and PTR have a shorter convergence time than Pathload.

30. Presentation Date : 7 Sep 200530 Conclusion The probing packet size and packet train length are investigated to have a achieve the best measurement accuracy. IGI method loses accuracy if the tight link is not the bottleneck link, or if there is significant competing traffic on links following the tight link.