One Rapid Flow With ON/OFF UDP Traffic

One rapid from smithfd-rapid22 to smithfd-rapid21, Gap Epsilon= 2us, Alpha= 0.875/0.9375 for exponential smoothing. The single rapid flow lasts for 1 minute. UDP Cross traffic from smithfd-rapid6 to smithfd-rapid3 lasts from 20s to 40s.

Throughputs:

a. With 300Mbps CBR



b. With 500Mbps CBR



NOTE: The reason why gap=1us fails to work is simply because that the excursion threshold is too small. When alpha= 15/16, we smooth too much and the result becomes even worse. I verified abests with my scripts for the experiment and the calculation was correct. But when cross traffic enter, 1us becomes a better choice.

Examples:

In pstream 120, the first large value contributes to the single excursion starts from the very begining. Abest= 53Mbps.

In pstream 756, 1us is too small and excursion from 0-13, abest= 81Mbps.

Analysis of ABests with logs from the experiment with gap_ns_epsilon= 2us and alpha=0.875 with 300Mbps CBR only



The plot skips those probe streams which are in loss recovery phase and only shows the probe streams with ab\_estimations continously. The ab_estimations all fill in the range of actual_ab*m and actual_ab/m m=1.15. I do not know what we can do to improve.

One Example of Underestimation: pstream 2154 ABest=222Mbps

pstream 2160 ABest=80Mbps

One Example of Overestimation: pstream 2234 ABest=762Mbps

pstream 1165 ABest= 746Mbps

Sendrate:



Number of segments dropped: around 110 for another experiment with the same setting

Analysis of ABests with no cross traffic

99% ab\_estimations are larger than 880Mbps.

One Example of Underestimation: pstream 1969 ABest=512Mbps

The first spike leads to low abest

pstream 1975 ABest=492Mbps

pstream 1570 ABest=258Mbps

pstream 1580 ABest=379Mbps

EXP_AVG leads to low highest smoothed probing rate