COSMOS: A Simulation Framework for Swarm-Based Orchestration in the Edge-Fog-Cloud Continuum

Introduction

Here we present COSMOS (Continuum Optimization for Swarm-based Multitier Orchestration System), a novel simulation framework for evaluating self-organizing scheduling algorithms in edge-fog-cloud ecosystems. Unlike conventional schedulers that rely on centralized optimization, COSMOS implements swarm intelligence principles inspired by decentralized biological systems, enabling emergent coordination across distributed nodes.

The framework extends Mesa’s agent-based modeling toolkit to simulate:

• Multi-tier resource dynamics: Agent populations representing edge devices, fog nodes, and cloud servers with configurable behavioral policies
• Constraint-aware scheduling: Integration of answer set programming (ASP) for hard constraints and metaheuristic optimization for soft constraints
• Network-aware orchestration: Evaluation of communication costs and dependencies across continuum layers

Developed in Lakeside Labs in scope of Myrtus (Multi-layer 360 dynamic orchestration and interoperable design environment for compute-continuum systems) project.

Installation

Make sure the right Python version is installed

Make sure that python binary is of the correct version. For Windows users, Python 3.10 can be installed from Microsoft Store. Linux users already know how to install Python on their system. The simulation framework was tested with Python 3.10 and Mesa 3.0.3 (with corresponding dependencies). To check the version of your binary, execute the following command:

python --version

Dependencies

The following Python modules are required for this project. Ensure that you have the specified versions installed:

• mesa (3.0.3)
• matplotlib (3.5.1)
• numpy (1.21.5)
• networkx (3.3)
• protobuf (3.12.4)
• pyyaml (6.0)

Additional Dependencies (installed via mesa)

• pandas (2.0.3)
• tqdm (4.66.4)
• python-dateutil (2.8.2)

To install the required dependencies, run:

python3.10 -m pip install -r requirements.txt

Or:

python3.10 -m pip install mesa==3.0.3 matplotlib==3.5.1 numpy==1.21.5 networkx==3.3 protobuf==3.12.4 pyyaml==6.0

On Windows

In case of issues with installing pyyaml, the following steps can be taken:

python3.10 -m pip install "cython<3.0.0"
# then run:
pip install --no-build-isolation pyyaml==6.0

Proceed further with the installation of the remaining dependencies from requirements.txt.

Running the simulation

To run the simulation:

• adjust the configuration file named config.yml in the config directory
• navigate to the modules directory:

cd modules

• execute the following command:

python3.10 EFC_main.py

Directory Structure

The directory structure of the simulation framework is as follows:

• modules: contains the source code structured in modules.
  • config: contains the configuration file.
  • generated/data contains the generated dataset of the EFC network graphs/models.
  • trained_mlp contains the configuration files of the trained MLP models.
  • interfaces contains compiled protocol buffer files.
  • evaluation contains the script eval.py for evaluation of the trained MLP models.
• proto: contains the protocol buffer files describing the inter-process communication with the FREVO training framework. Is not needed for running the simulation.

Configuration

The configuration file config.yml contains the following parameters:

• num_pods: number of pods in the network. Must be bigger or equal to sim_steps.
• sim_steps: number of simulation steps. The number of pods entering the network at each step is calculated as: num_pods / sim_steps
• num_edge_devices: number of edge devices in the network. The num_ parameters are not considered when the graph is imported with import_graph: True.
• num_fog_nodes: number of fog nodes in the network.
• num_cloud_servers: number of cloud servers in the network.
• behaviour_type: type of decision-making behaviour ('random_walk', 'aco' or 'mlp'). The 'mlp' behaviour requires a trained MLP model to be imported if mode: normal is set.
• import_graph: whether to import a graph from the generated/data directory.
• graph_sample: name of the graph to import. The graph must be in the generated/data directory. For example, if the file is named EFC1.graphml, the value of graph_sample should be EFC1. The right format is .graphml.
• nn_sample: name of the MLP configuration file to import. The file must be in the trained_mlp directory. For example, if the file is named mlp_config.yml, the value of nn_sample should be mlp_config.yml.
• plot_all: whether to plot the simulation results. Is disabled during the training phase.
• gen_samples_num: number of samples to generate. Is only used when mode: gen is set.
• mode: mode of the simulation ('normal', 'gen' or 'train'). The 'train' mode trains an MLP model using the FREVO training framework. The 'gen' mode generates a dataset of EFC network graphs/models.
• path_to_dataset: path to the dataset. Is only used when mode: gen is set or when the graph is imported. The following parameters are only used when behaviour_type: mlp and mode: normal are set:
• frevo_port: port number of the FREVO training framework to connect to.
• candidate_evals: number of inference requests to evaluate a candidate.
• training_results_path: path to save the best trained model.
• evaluate_best_candidate: whether to evaluate the best trained model at the end of training+exporting.
• randomized_pod_interval: whether to randomize the interval between the pods entering the network. If set to False, the interval is equal to 1 and number of issued pods is always predictable.
• mu: parameter µ ∈ (0, 1] defining the frequency of randomly arriving pods. The value is only used when randomized_pod_interval: True is set.

Typical configuration:

num_pods: 500
sim_steps: 10
num_edge_devices: 35
num_fog_nodes: 10
num_cloud_servers: 5
behaviour_type: 'aco'
import_graph: True
graph_sample: EFC1
nn_sample: 'mlp_config.yml'
plot_all: True
gen_samples_num: 50
mode: 'normal'
path_to_dataset: 'generated/dataset'
frevo_port: 9990
candidate_evals: 10
training_results_path: "/home/user/Documents/Myrtus/Frevo/Frevo/Results/EFC_test"
evaluate_best_candidate: True
randomized_pod_interval: False
mu: 0.5

Automatic evaluation

Saved trained MLP candidates can be evaluated automatically using the combination of EFC_Evaluator.py source file located in modules directory and eval.py script located in the modules/evaluation directory. To generate the evaluation results, the following steps should be taken:

• adjust the configuration file named config.yml in the config directory
• navigate to the modules directory:

cd modules

• run EFC_Evaluator.py:

python3.10 EFC_Evaluator.py

The preliminary evaluation results for plotting will be saved in the evaluation directory as eval.csv.

• navigate to the modules/evaluation directory:

cd evaluation

• run eval.py to visualize and plot the evaluation results:

python3.10 eval.py