Objectives

Motivation:  Knowledge of CWND and RTT helps determine many important characteristics of sender, receiver, and the network path. For example:



Methodology


Basic Idea:   Construct a "replica" of the TCP sender's state for each TCP connection observed at the measurement point. Emulate state transitions:


Challenges:


All of these introduce uncertainties into CWND estimation.





Ideas used in meeting the challenges:





RTT estimation:






Sources of Estimation Uncertainty


Under-estimation of CWND:


Over-estimation of CWND:



Window-scaling:



TCP Implementation Issues:





Validation


Simulations:
Figure 2




Internet Experiments:
Figure 3



Conclusion:

Relative errors were small, often of the order (and possibly due to) rounding errors or clock tick precisions.





Backbone Measurement Results


Trace Characteristics:

Congestion Window:



TCP Flavors:



Greedy Senders:



Round-trip Times:



Efficiency of Slow-Start:





Comments:


Hope for the future:

ECN senders explicitly encode the CWND in packet header (for receivers). Monitors could just read this value off the packet headers. Currently, ECN not widely deployed, though (< 0.14%).