Knowledge-Defined Networking Training Datasets

This website aims to publish training datasets to encourage open research, development and benchmarking of Machine Learning algorithms applied to Computer Networks.


Network modeling datasets

Datasets v0

Dataset description

NSFNet Topology

Name Nodes Size MD5Sum
nsfnet.tar.gz 14 2.0 GB d1c2bf8d461b571f413f9df4ec845c33

GEANT2 Topology

Name Nodes Size MD5Sum
geant2.tar.gz 24 6 GB 4847eb2c4c14c6c2e1ba82287ffc7365

50 Nodes Topology

Name Nodes Size MD5Sum
synth50.tar.gz 50 29.0 GB 993e8cbe8f897235add3ae9aa1c47a82

Datasets v1

Dataset description

NSFNet Topology

Name Nodes Size MD5Sum
nsfnetbw.tar.gz 14 1.1 GB 5d800b709721925fcc291b9e4eb2da11

GBN Topology

Name Nodes Size MD5Sum
gbnbw.tar.gz 17 1.7 GB efe9ce8aabda216e93b3bc41c44cbc22

GEANT2 Topology

Name Nodes Size MD5Sum
geant2bw.tar.gz 24 3.3 GB 51a995fd6e461cbb8008407da44f8737

Germany 50 Nodes Topology

Name Nodes Size MD5Sum
germany50bw.tar.gz 50 15.8 GB b738280819bccfadcca1be57085bc69e

Datasets papers

Virtual Network Functions (VNFs):

Name Size Description Includes
OVS.zip 527 kB CPU consumption of an OVS connected to a SDN controller. See further details in the KDN paper Dataset (traffic features + cpu consuption), scripts to read the dataset, readme.txt
Firewall.zip 530 kB CPU consumption of an OVS configured with firewall rules. See further details in the KDN paper Dataset (traffic features + cpu consuption), scripts to read the dataset, readme.txt
Snort.zip 532 kB CPU consumption of a SNORT with the initial configuration. See further details in the KDN paper Dataset (traffic features + cpu consuption), scripts to read the dataset, readme.txt

Network Characterization:

Name Size Description Includes
overUnderNetwork.zip 17.35 MB Flow-level delay in a overlay-underlay network with load balancing in the overlay Dataset (traffic + load balancing + delay), readme.txt
star.zip 670.8 MB Delay among pairs of nodes in a star network Dataset (traffic + delay), readme.txt
ring.zip 671.6 MB Delay among pairs of nodes in a ring network Dataset (traffic + delay), readme.txt
scaleFree.zip 911.8 MB Delay among pairs of nodes in a scale free network Dataset (traffic + delay), readme.txt
overUnder.zip 907.1 MB Delay among pairs of nodes in a overlay-underlay network Dataset (traffic + delay), readme.txt

Unveiling the potential of GNN for network modeling and optimization in SDN: Paper

Name Size Description
nsfnet.zip 1.3 GB Datasets of the NFSNet topology (This dataset was randomly partitioned to obtain the training and evaluation datasets)
geant2.zip 5.2 GB Datasets of GEANT2 topology (Used only for evaluation)
GBN.zip 2.8 GB Dataset of the GBN topology (Used only for evaluation)
trained_model_delay.zip 3.0 MB RouteNet model for delay trained with the NSFNet topology dataset
trained_model_jitter.zip 1.0 MB RouteNet model for jitter trained with the NSFNet topology dataset (This model was trained from a model of the delay in an early training stage. See further details in the paper)
README_gnn.pdf 35 KB Description of the datasets

Understanding the Network Modeling of Computer Networks:

Name Size Description Includes
routing.zip 120.3 MB Flow-level delay in a scale-free network with 4 different routings Dataset (traffic + delay)
saturation.zip 1.4 GB Delay among pairs of nodes a 10-nodes scale-free network changing the traffic intensity and traffic distribution Dataset (traffic + delay), readme.txt
netSize.zip 46.0 MB Delay among pairs of nodes in a 5, 10 and 15 nodes scale-free network. Dataset (traffic + delay)
topologies.zip 1.3 GB Delay among pairs of nodes different topologies and sizes. Dataset (traffic + delay), readme.txt

Neural Nets:

Name Size Description Includes
codeMatlab.zip 916 kB Matlab code used to train the VNF data and the figures from the experiment described in the KDN paper Code, readme.txt
codePython.zip 6 kB Python code used to train the overUnderNetwork data Code, readme.txt

Ready to use examples:

Name Size Description Includes
KDN_overlayUnderlay.zip 18,2 MB Code, dataset and setup fot the overlay-underlay experiment described in KDN paper Ready to set up and run

A Deep-Reinforcement Learning Approach for Software-Defined Networking Routing Optimization

Name Size Description Includes
DRL.zip 43 KB Source code for the Deep-Reinforcement Learning agent and OMNeT++ network topology Code, readme.md
train.csv.gz 34 KB Evolution of DRL agent's performance during training, measured in average network delay Dataset
benchmark.csv.gz 482 MB Performance benchmark using 1000000 routings times 1000 traffic matrices, measured in average network delay Dataset
traffic.zip 587 KB 1000 test traffic matrixes (100 per each traffic intensity) Dataset
100k.csv.gz 4,3 MB 100000 test routing configurations Dataset


Papers:


Contact:

Site maintained by Unviersitat Politècnica de Catalunya, feel free to contribute with your training sets by sending an email to:

Albert Cabellos: