This repository holds the implementation of FutureCoord, presented in our paper "Use What You Know: Network and Service Coordination Beyond Certainty" (author version) accepted at 2022 IEEE/IFIP Network Operations and Management Symposium. FutureCoord combines Monte Carlo Tree Search with Traffic Forecasts for Online Orchestration of Network Services.

Our paper considers a network optimization problem where an orchestrator must optimize the placement of single virtual network functions (VNFs) as part of service function chains (SFCs). An assignment is only successful if all VNFs of a SFC are assigned sufficient amounts of resources (memory, cpu, ...) and the quality of service (QoS) constraints (max. latency, ... ) are not violated. The goal of our orchestrator is then to find an allocation of resources that successfuly deploys as many SFCs as possible on the network while the incoming traffic and their requested SFCs continously change over time.

We propose an algorithm that optimizes the discrete decision problem according to the predictions of a traffic model. Said model forecasts what SFCs will continue to be in demand and what load is to be expected. The optimization algorithm can then decide to allocate more resources to VNFs that are more in demand or will be in the future.

If you use this code, please cite our paper (author version):

@inproceedings{werner2022futurecoord,
	title={Use What You Know: Network and Service Coordination Beyond Certainty},
	author={Werner, Stefan and Schneider, Stefan and Karl, Holger},
	booktitle={IEEE/IFIP Network Operations and Management Symposium (NOMS)},
	year={2022},
	publisher={IEEE/IFIP}
}

Assuming an Anaconda distribution has already been installed, the environment can simply be cloned via conda env create -f environment.yml. We tested this setup on an Ubuntu 18.04 machine with Intel Xeon [email protected] CPUs and 64GB RAM.

The script.py file serves as an interface to running any experiment. It can be used to specify what experiment to run, what service coordinator to use, how many episodes to evaluate and where evaluation files are saved. The interface is as follows:

python script.py
  --experiment data/experiments/<topology>/trace.yml
 --agent data/configurations/<coordinator>.yml 
 --logdir <path to logdir>
 --episodes <#evaluation episodes>
 --seed <random seed>

The experiment and algorithm configurations are saves as *.yml files under data/experiments and data/configurations, respectively. Each experiment logs summary files of its evaluation, i.e. experiment, algorithm, etc., as summary.yml in the <logdir>. Monitoring information including the obtained completion ratio as well as resource utilizations are logged as results.csv.

In our evaluation, we vary the expected data rate, max. delay bound and arrival rates for all kinds of flows via the scenarios/flows.py file. It can be called as follows:

python scenarios/flows.py 
  --experiment data/experiments/<topology>/trace.yml
  --agent data/configurations/<coordinator>.yml
  --logdir <path to logdir>
  --episodes <#evaluation episodes>
  --seed <random seed>
  --property <datarate / latency / load>
  --factors <varies mean of distributions>
  --sim_factors <varies mean of distributions for forecasts>
  --traffic <accurate / erroneous> 
  --pool <number of jobs>

The parameter --factors varies the flow distributions' mean values proportionally either in terms of their latency, data rates or arrival rates (--property). Similarly, parameter --sim_factors varies the forecast distributions' mean values. Whether or not the forecast flows follow the correct pattern of arrival rates (or another episode's) is decided by the --traffic parameter. Similarly, the files scenarios/network.py and scenarios/searches.py define interfaces to vary either the compute and link capacities or the number of search iterations performed by FutureCoord.