mmwave_cc

Testbed for measuring and improving TCP performance over mmWave links

Requirements

Python 3, NumPy, Pandas, Matplotlib, Make, Mahimahi.

Steps

• Build the client and server programs inside sender_receiver by typing make

• (optional) Test the client and server programs separately by running sender_receiver/sender and sender_receiver/receiver in two separate shell instances. The behavior is similar to the iperf utility, and the options are a subset of the options available to iperf.

  • receiver needs a port, max number of connections before it exits, a verbose option and optional file name for logging the output.

  • sender needs the server IP address and port. Options are sending mode, block size (default: 128 KiB), congestion control algorithm to use (default: system default) and optional file name for logging. Block size is the amount of data sent in a single send() call. Sending mode is one of time to run (-t) in seconds, number of blocks (-b) or trace file (-f). Trace file is expected to contain one number per line showing the millisecond. The number of occurrences of a millisecond number indicates the number of packets to be sent in that single millisecond. The CC algorithm must be one of the available ones in /proc/sys/net/ipv4/tcp_available_congestion_control.

• Run runmm.py with mandatory channel trace argument (only the name, not the full path). Channel traces are placed in traces/channels. More channel traces may be added to that directory. This program starts a Mahimahi mm-link shell to emulate the channel, runs the sender and receiver in separate subprocesses, saves the output trace and makes a plot of throughput and delay. Options to runmm.py include those for the sender and receiver program and the following: buffer length in bytes for Mahimahi (default: infinite), mm side (default: "receiver"), whether to use the system iperf utility (--iperf) and whether to display/save the plots. The mm side argument should be one of "sender" or "receiver". If it is "receiver", the receiver program is run inside the mm-link shell, else the sender is run inside mm-link. Note that we have observed problems with running the sender inside mm-link for TCP flows so we recommend sticking to the default "receiver" option for now. The --iperf option is currently not implemented.