Skip to content

Every Node for Itself: Fully Distributed Service Coordination

Notifications You must be signed in to change notification settings

CN-UPB/distributed-coordination

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Fully Distributed Service Coordination

Two fully distributed algorithms for online service coordination. All nodes runs the algorithm individually and in parallel to decide how to scale and where to place service components as well as how to route incoming flows through the placed instances.

This fully distributed approach reaches similar solution quality as centralized approaches but is more robust, requires less global knowledge, and is magnitudes faster.

This repository contains prototype implementations of both algorithms, extends coord-sim for simulation, and contains extensive evaluation results.

Citation

If you use this code, please cite our paper:

@inproceedings{schneider2020distributed,
	title={Every Node for Itself: Fully Distributed Service Coordination},
	author={Schneider, Stefan and Klenner, Lars Dietrich and Karl, Holger},
	booktitle={International Conference on Network and Service Management (CNSM)},
	year={2020},
	publisher={IFIP/IEEE}
}

Setup

Requires Python 3.6+. All dependencies can be installed with:

python setup.py install

For evaluation, also install these dependencies:

pip install -r eval_requirements.txt

Usage

  • All relevant scripts for running experiments are in src/algorithms/execution
  • The folders time and 3x3 are for different experiments but contain similar scripts. Choose either one or create a new one.
  • Create configurtion files by running config_creator.py. The generated configurations specify percentage of ignress nodes, capacities, etc.
  • Adjust the network, algorithms, ingress in iterator.py
  • Then call python iterator.py start_run end_run num_parallel poll_pause to run experiments.
  • For example python iterator.py 50 55 4 5 will run 6 repetitions (ID 50-55) on 4 cores, polling every 5s if a core is free.
  • The results are saved in the subfolder scenarios according to run ID, config, algorithm, etc.
  • Attention: python iterator.py silently overwrites existing results!

Evaluation of Results

  • To evaluate the results, aggregate and plot them.
  • To aggregate, run aggregator.py. You can adjust settings in settings.py.
  • In case of 3x3, also specify the runs to aggregate, eg, python aggregator.py 0 49.
  • Then plot with plotter.py. Or easier using/extending the available Jupyter notebooks under execution: 3x3/eval.ipynb, 3x3/eval_decisions.ipynb, time/eval.ipynb.
  • These notebooks use and plot the available evaluation results, which are stored in the scenarios and transformed folders.

Simulation

For simulation, the coord-sim simulator was extended as follows:

  • Forwarding capabilities. Prior to this, forwarding happened implicit. Now flows are explicit forwarded over link, taking into account individual link utilization
  • Individual flow forwarding rules. A node can make a forwarding decision on individual flows.
  • Individual flow processing rules. A node can explicit decide if it will process a flow.
  • Extended simulator state and action interface.
  • Algorithm callback interface. To allow a external algorithm to capture certain event, it can register callback functions invoked by the flowsimulator.
    • init_flow callback
    • pass_flow callback
    • periodic measurement callbacks.
  • Adapted metrics.
  • Egress node routing.
  • Extended simulator configuration

Contributors

Please use GitHub's issue system to file bugs or ask questions.