Parameter Tuning For Good AB_est

New Plots 700 UDP
Paramters: Size of Averaing, TINY_EXCURSION, GAP_NS_EPSILON

Description: Get CDF of udp throughput of each sliding window. We expect that the cdf of ab_ests should match the cdf of udp since they are calculated in the same time scale.

1. No Averaging













2. Averaging Among 2 Acked Segments













3. Averaging Among 4 Acked Segments













4. Averaging Among 8 Acked Segments













5. Averaging Among 16 Acked Segments













Promising Parameter Sets: averaging_size=4/8, excursion gap threshold=1500,3000

1. CDF of plot of the ratio between abest and ab for each excursion









2. CDF of plot of AB_est and actual AB throughout the whole pstream









3. Pstreams with Excursions Starting at probing rate < 500Mbps

(1) Probe Streams Underestimating AB_est





When gap threshold is changed to 3000 or valid excursion to 2, then the first excursion will not be detected. When gap threshold is still 1500, and valid_excursion=2, three short excursions(probe 5-6, probe 10-13 and probe 28-end) will be detected which deprecates the AB_est.





If gap threshold is increased to 3000, or valid excursion decreased to 2, then the only excursion detected starts from the exact segment when probing rate + udp rate began to exceed 1Gbps.



To summerize the results for all sets of parameters for pid=11865:

average udp throughput throughout the whole pstream: 317Mbps
average ab: 683Mbps
avg size	valid ex	gap threshold	result		
4		1		1500		527
4		2		1500		541
4		4		1500		585
4		1		3000		552
4		2		3000		561
4		4		3000		605
8		1		1500		517
8		2		1500		517
8		4		1500		517
8		1		3000		706
8		2		3000		706
8		4		3000		706	

So for some probe streams which under-estimate available bandwidth, we should avoid detecting excursions when probing rate is relatively low. For approporiate size of averaging, the burstiness of recv gaps is removed. The inspirations from these experiments are:


(1) averaging is required to remove the burstiness nature of send gaps and recve gaps
    averaging_size = 4/8 is promising. The smaller the averaging size is, the finer granularity we can get about AB_est. However, the larger the averaging size is, the more accurate abests are. So here is a trade-off.
(2) valid excursion size must be larger than 1. With sliding window averaging, we can get many data points. 
    So valid_excursion = 2/4 can ensure greate granularity. valid_excursion cannot be too large to infate abest.
(3) gap_threshold must be carefully chosen. Some value between 1500 to 3000 should work well for this experiment.
(4) early detecting of excursions when probing rate is relatively low should be avoided.
For valid excursions, how about requiring q_max to be at least a certian proportioin of send gap? For example:
 if qmax < 0.05*largest_send_gap_in_that_excursion, then this excursion should not be counted as a valid excursion. 

(b) Probe Streams Over-estimating AB_est







Over-estimation is mainly caused by large gap threshold values.

When UDP Cross Traffic is 700Mbps