Ns3-fl is a federated learning simulator which takes into consideration data, algorithm, and network. Ns3-fl is built off of two existing simulators, flsim and ns-3. Flsim is the federated learning simulator, which simulates data and algorithm over an ideal network. Ns-3 is an open source discrete event network simulator, created for research and educational use. Ns3-fl also contains an power module which computes the computational time, power, and energy, as well as the transmit power, and energy.
To install ns3-fl, first clone the respository. To clone the submodules, run
cd ns3-fl
git submodule update --init --recursiveNext, install Anaconda.
To configure the ns3-fl-network submodule, from the root directiory, run
cd ns3-fl-network
./waf configure
./wafTo configure the flsim submodule, from the root directory, run
cd flsim
conda env create -f environment.yml
conda activate fl-py37The above commands create and activate an Anaconda enviroment, with all of the dependencies needed to run flsim. For a more detailed description on setting up and running flsim, refer to the original flsim repository.
To start a simulation, from the root directory, run
cd flsim
python3 run.py --config=configs/config.json To specify simulation parameters, we use the same configuration JSON files as in flsim, though we add more fields to allow for a wider range of data and algorithm settings, in addition to network settings.
There are examples of how to set up the configuration file in flsim/configs/ns3-fl.
The configuration options which are not listed below were a part of the original flsim. Refer to the original flsim wiki for details on how to set these parameters.
In the nested model section model:
size: Specifies the model size in kB. The model size will be used in the network code.
In the nested section network, folder: Specifies folder in the scratch file which should be run for the network. The default parameter, which works with the rest of our network settings, is wifi_exp. type: Specifies the type of network to run the simulation on. Currently supported options are ethernet and wifi. The network code is in the scratch folder of the ns3-fl-network submodule. wifi: A nested section, within the network section, that specifies configurations for the WiFi network.
loss: Specifies the mean of the range of propogation loss in the network. We use the RandomPropogationLossModel in the network code to simulate loss. The range for the loss is 20 below and above the specified mean. max_packet_size: Specifies the maximum size for a packet to be sent over the network. ethernet: A nested section, within the network section, that specifies configurations for the Ethernet network.
max_packet_size: Specifies the maximum size for a packet to be sent over the network. device_type: Specifies the device type for the clients in the network simulation. The device type is used when calculating the average power and energy consumed during each round of the simulation. Currently, the supported options are 400 and 4 for Raspberry Pi 400s and 4s, respectively.In the event that the submodules are not directly accessible, here are the links to the submodules:
These repos can be seperated cloned into the ns3-fl folder. The configuration process will be the same as above.