1. Development of a New Efficient and Accurate Available Bandwidth Estimation Method Péter Hága Attila Pásztor István Csabai Darryl Veitch Viktória Hunyadi

2. CNL - Network Performance Measurement Group 2 Outline Available bandwidth measurements, motivations Definition of Available Bandwidth Milestones in the evolution of AB measurement Measurement Method, and results from simulation Analytic Approximation Estimation of AB Laboratory Experiments Conclusion and Future Plans

3. CNL - Network Performance Measurement Group 3 Importance of available bandwidth measurements, motivations The knowledge of available bandwidth along a route is important: for call admission control congestion control and avoidance determine the instantaneous free transmission capacity The motivations behind developing methods to estimate AB: applications cloud use it to adopt their behaviour avoid congested paths determine the optimal transfer rate without losing packets

4. CNL - Network Performance Measurement Group 4 Definition of Available Bandwidth Available bandwidth of hop h, over the time interval  T,T  (roughly the difference of the link capacity and the average cross traffic) Available bandwidth of a route is a smallest hop AB on a path:

5. CNL - Network Performance Measurement Group 5 Milestones in the evolution of AB measurements problems, advantages Pathload: binary search algorithm invasiveness - high rate probing results burst of packet trains time - iteration requests significant amount of time non-stationary of cross traffic - binary search can results contradictory results PathChirp: chirp based measurement estimation from measured data is heuristic fine timescale New class of probing techniques: interaction class (between probes and CT) not requiring a knowledge of the link bandwidths not requiring a knowledge which link is a narrow link

6. CNL - Network Performance Measurement Group 6 Our goal improve a method to estimate AB, which: based on theoretical model can give us accurate estimations with high frequency develop a tool to make our accurate estimations in robust way make this tool widely accessible for use it in smart applications (eg. transfer protocols)

7. CNL - Network Performance Measurement Group 7 Measurement Method Sending probes with increasing rate (like pathChirp) Measuring the one-way delay : as we reach the free capacity of a link the OWD increases sharply

8. CNL - Network Performance Measurement Group 8 Measurement Method Instead of measuring OWD we will measure the ratio of inter arrival time and the inter probe spacing (t*/t) to avoid problem comes from the unsynchronized clocks of the sender and receiver t  t , as LPR  (we are below the AB) t*/t  linear, as LPR  available bandwidth (we above the AB)

9. CNL - Network Performance Measurement Group 9 Measurement Method t*/t is sensitive for the mean of the free capacity along a route, and also for the coefficient of variation (  ) of the CT

10. CNL - Network Performance Measurement Group 10 difficulties in real measurements: fluctuations in the measured data ( even in stationary CT )  averaging over some chirps small number of probes  not to cause serious congestion sparsely spaced probes to cover the whole regions of the possible free capacity sparsely spaced probes  we need a function to fill out the space between two measured points Measurement Method

11. CNL - Network Performance Measurement Group 11 f  x , as probing rate x  smooth and monotone increasing asymptotically linear for large x  f  x  x  The function we are looking for: The function has at least 3 parameters: two to fit the asymptotic slopes one to allow for the Cross Traffic dependent speed of transition Analytic Approximation

12. CNL - Network Performance Measurement Group 12 Analytic Approximation The function is: We can determine the parameters of best fit using the Levenberg-Marquardt Algorithm.

13. CNL - Network Performance Measurement Group 13 Analytic Approximation The fits were generally good both for Poisson and Weibullian-renweal cross traffic (with different mean and coefficient of variation (  ))

14. CNL - Network Performance Measurement Group 14 Estimation of Available Bandwidth  th  x  where f  x   int  c  log  a) Two algorithms to select the transition scale: threshold based:  th  x  f  x  s  (we used s  intersection based ( the intersection of the two asymptotes ):  int 

15. CNL - Network Performance Measurement Group 15 Comparing the performance of  th  and  int  th : sensitive to mean, but worsen as burstiness increases  int : also sensitive to burstiness, Estimation of Available Bandwidth

16. CNL - Network Performance Measurement Group 16 Laboratory Experiments Validation: compare with the standard pathChirp tool in controlled environment Network testbed: Probe stream Controlled CT - Receiver tool - DAG monitor - CT generator - Sender tool

17. CNL - Network Performance Measurement Group 17 Laboratory Experiments estimations by pathChirp, and by the combinations of the intersection and threshold methods

18. CNL - Network Performance Measurement Group 18 Conclusions and Future plans we have developed an improved method for available bandwidth estimation, resulting better estimations first steps to work out the underlying theoretical model for evaluation of the measured data In the future: develop a complete measurement package explore different strategies for sampling the investigated curve find more robust alternative transition scale selection algorithms

19. CNL - Network Performance Measurement Group 19 Thank you for your attention!