Since there are thousands of different flows, should we select a few flows to analyze some of the per-flow statistics (such as flow duration and flow size)?
You are drawing the CDF. Each flow gives you one sample (e.g., one duration value). Then you draw the distribution of these values (e.g., flow duration) as one CDF.
Just to confirm, for the flow inter packet arrival time cdf, one flow having n number of packets will give n samples to be used in the CDF correct?
So the total samples will be (number of flows) x (number of packets in each flow)?
Yes. That is true for inter arrival time CDF.