1. "Coded probes," or a group of temporally structured packets that are robust against random queuing delays and timestamping noise. The paper proposes that the sender sends two packets with a fixed time interval so if the receiver gets some other interval this data pair must have been corrupted. I didn't find any argument on why the scheme is good at identifying noise points besides a note about a 4-5x improvement. I don't think this scheme necessarily identifies all noise points but it should be able to eliminate some of the true positive noise points because of the temporal constraint. More temporally sophisicated structures may have even better performance at this task.
2. So they collected a bunch of the coded probes and plotted them and found out it somehow looked like the classic plot of the linear SVM, or actually the error model of the measurement scheme being a stable lower bound of transmission time plus a positive somewhat exponentially distributed error. They use a SVM to identify the lower bound because the data fits the SVM perfectly. I think more sophisticated models can also apply here given proper probabilistic accounting of the error model.
3. Cooperative network synchronization. In essence getting better accuracy by averaging several clocks but in a network you can't average them directly so it becomes message-passing on a probabilistic graphical model. As an analogy think of a network of springs, and the network is optimized/converged when all springs are at their lowest energy levels.
Why do you find that to be substantial? Given most machines are at least 50% idle, I find 0.5% to be acceptable.
will show if you have a PTP support for your NIC.