The goal of the exercise is to determine the impact of the warm-up time on the results when performing steady-state simulations. We will be working with ms-lab6.cc which is based on examples/tcp/tcp-bulk-send.cc. As a performance measurement we will be evaluating the average throughput.
Create a copy of ms-lab6.cc in ns-3's scratch directory. Run the simulation with RngRun set to your index/album number. Open one of the generated .pcap files in Wireshark and select Statistics->IO Graph from the menu.
Set the interval (for displaying the results) to 10 ms and no smoothing. At what time does the warm-up period end? Provide a screenshot to support your claim.
Experiment with the smoothing options by selecting various moving average filters. Which moving average range does, in your opinion, make the transient behaviour easiest to observe? Which moving average range is too large and hides the warm-up period? How about if we changed the interval to a larger value?
To obtain a benchmark (reference result) we need to calculate the average throughput over a long simulation run. Why will the impact of the transient period be negligible?
To achieve this, set the simulation time to 500 seconds in your copy of ms-lab6.cc.
After running the simulation, use Wireshark to calculate the average throughput:
ip.src==10.1.1.1,What is the average Mb/s of displayed packets?
Now set the simulation time to 25 seconds and calculate the average throughput for various warm-up times (between 0 and 20 seconds) for a single simulation run. You can implement the warm-up time in Wireshark by adding frame.time_relative>x (where x is the warmup time) to the display filter (together with the ip.src filter). For each result calculate the following metric: the relative error between the obtained throughput values and the benchmark from Task 2.
Plot the calculated metric as a function of the warm-up time. Is this a monotonic function?