Joint Service Deployment - Manager Placement

Introduction

JSD-MP solves the joint placement of Virtual Network Functions (VNFs) and the VNF Managers (VNFMs) that manage them. Most placement work optimizes VNFs first and places managers afterward; we instead place both together, so VNF placement is steered toward layouts a manager can actually serve — subject to manager capacity, a latency radius, and co-location (notManagerNodes) constraints. The problem is NP-hard, so we provide heuristics and compare them against an exact CPLEX solver.

Central finding. Joint placement is best understood as a feasibility- preservation mechanism, not a revenue tweak: a management-blind (disjoint) pipeline frequently strands whole chains because no admissible manager exists for the layout it chose. Its benefit is therefore conditional — large exactly when management constraints bind, and negligible (even slightly negative) when they do not. See Key findings and analysis/FRAMING.md.

Methods implemented here:

• bari — the joint VNF+VNFM heuristic (management-aware DP / Viterbi); the core method
• rari — MASMAN, a randomized variant
• abu — Abu-Lebdeh's tabu-search baseline, unmodified
• oabu — Abu-Lebdeh adapted to our constraints for a fair comparison
• Optimal with CPLEX, implemented under simulation/ (a Java/Gradle project)

In the jsd-mp problem some functions can be managened and there are the functions that need VNFM, so the following code snippet find these functions in a chain.

itertools.compress(
        placement.nodes, chain.manageable_functions
)

Also we have the following configuration for our VNFM constraints as default:

ram: 4
cores: 2
capacity: 10
radius: 100
bandwidth: 10
licenseFee: 100

Abu Method

Abu-Lebdeh describe a method based on tabu-search to improve VNFM placement on datacenter that already has VNF placement. In our work we are considering the VNF placement jointly with VNFM placement so we are going to change it to consider the placement and after that compare it with our method. In Abu-Lebdeh method there is no way to discard a chain so it can generate infeasible results so we are reserving resources for VNFM to prevent the infeasible situation.

The following variables are available in the src/jsd_mp/abu solution to configure it so you need to change them by hand (with --options) and report them into results.

n_iter
reserve_percentage

Optimized Abu Method

Abu-Lebdeh uses tabu-search to improve its VNFM placement but at this method it doesn't use consider our problem constraints, so here we are going to optimize it based on our constraints.

Key findings

From the re-analysis in analysis/, which reports chain acceptance rate (feasibility) with paired significance tests plus a controlled sensitivity sweep. Full detail in analysis/SUMMARY.md and analysis/SWEEP.md.

• Joint placement preserves feasibility. It accepts ~100% of chains across regimes, whereas disjoint placement is bimodal — a run either nearly succeeds or collapses (often to ~30–50% acceptance) because the post-hoc manager step is infeasible. Mean-revenue plots hide this; acceptance rate exposes it.

• The advantage is conditional on management binding:

  Regime	Joint accept	Disjoint accept	Reading
  USnet, default (slack) management	~100%	~100%	joint adds nothing; revenue −0.2…−0.9% (p<0.01)
  FatTree, default VNFM	~100%	~20–87% (bimodal)	joint preserves feasibility
  USnet + 4-hop radius	~100%	~35–94%	tightening one constraint flips the regime

• The constraint is the cause. Toggling the notManagerNodes co-location constraint on/off makes the joint advantage appear (+12–19 pp acceptance) and disappear (0 pp), and the gap decays as load saturates raw compute. On the tested topology, manager capacity is not a binding lever.

Results

Running a solver writes a report.csv that the Jupyter notebooks under results/ load. The corrected, statistically-honest re-analysis and the sensitivity study live under analysis/ and are regenerable with uv run --group notebook python analysis/<script>.py.

Paper

The manuscript lives under paper/, written in LaTeX using the MDPI Future Internet class (vendored in paper/Definitions/). Build the PDF with make -C paper (or latexmk -pdf -bibtex paper/main.tex); CI also typesets it on every change and uploads the PDF as an artifact. See paper/README.md for details.

Topologies

There are two topology that are considering here, fattree and usnet. There is another project that generate networks with these two topology. The generated configuration that must be copied and used with this project as follow:

topology fattree -k 4
cp topology.yaml ../jsd-mp/config/topology.yml

Please note that you must also consider to store these configuration for future re-runs.

Chains

We are going to place SFCs on our network and there is project that generate chains. The generated chains chains can be copied and used with this project as follow:

chainer -n 100
cp chains.yaml ../jsd-mp/config/chains.yml

How to Run

This project is managed with uv. Install the dependencies (and the jsd-mp package itself) with:

uv sync

Then run a solver via the jsd-mp console entry point:

uv run jsd-mp -ss rari -c config/ -r 10

Run the test suite and type checker with:

uv run pytest
uv run mypy

The Jupyter notebooks and the analysis/ scripts need the extra notebook dependency group:

uv sync --group notebook
uv run --group notebook python -m ipykernel install --user --name=jsd-mp-krnl

In your notebook, Kernel -> Change Kernel. Your kernel should now be an option.