Page 17 - ITU Journal Future and evolving technologies Volume 2 (2021), Issue 4 – AI and machine learning solutions in 5G and future networks

ITU Journal on Future and Evolving Technologies, Volume 2 (2021), Issue 4

GRAPH‑NEURAL‑NETWORK‑BASED DELAY ESTIMATION
FOR COMMUNICATION NETWORKS
WITH HETEROGENEOUS SCHEDULING POLICIES

1
1
1,2
1
Martin Happ , Matthias Herlich , Christian Maier , Jia Lei Du , Peter Dor inger 1
1 Intelligent Connectivity, Salzburg Research, Austria, IDA LAB, University of Salzburg, Austria
2
NOTE: Corresponding authors: Martin Happ, martin.happ@sbg.ac.at; Peter Dor inger, peter.dor inger@salzburgresearch.at
Abstract – Modeling communication networks to predict performance such as delay and jitter is important for evaluat‑
ing and optimizing them. In recent years, neural networks have been used to do this, which may have advantages over
existing models, for example from queueing theory. One of these neural networks is RouteNet, which is based on graph
neural networks. However, it is based on simpli ied assumptions. One key simpli ication is the restriction to a single
scheduling policy, which describes how packets of different lows are prioritized for transmission. In this paper we pro‑
pose a solution that supports multiple scheduling policies (Strict Priority, De icit Round Robin, Weighted Fair Queueing)
and can handle mixed scheduling policies in a single communication network. Our solution is based on the RouteNet ar‑
chitecture as part of the ”Graph Neural Network Challenge”. We achieved a mean absolute percentage error under 1% with
our extended model on the evaluation data set from the challenge. This takes neural‑network‑based delay estimation one
step closer to practical use.
Keywords – Communication networks, delay estimation, graph neural networks, scheduling

1. INTRODUCTION simulators such as OMNeT++ [5]. However, such sim‑
ulations may be time‑consuming. If the communication
There has been an increasing use of machine learning network itself or any settings are changed, the simula‑
techniques for various kinds of problems in recent years. tion has to be repeated. Thus, it becomes especially time‑
Due to the variety of problems, many new machine learn‑ consuming when simulating the impact of a series of pa‑
ingalgorithmshavebeendeveloped. Inparticularfordata rameter changes. In contrast, the time‑consuming train‑
that can be described by graphs, there has been an im‑ ing of neural networks has to happen only once in general.
portant new development known as ”Graph Neural Net‑ Hence, prediction with neural networks usually provides
works” [1]. Examples of such data are chemical elements a faster way to estimate the performance of networks.
or communication networks. We focus on the latter in this
paper. In this context, a communication network can be 2. RELATED WORK
characterized by nodes and edges, where the edges rep‑
resent the links between nodes. Additionally, there are There are classical (i.e. non‑machine learning) methods
properties associated with each node and each edge. This to predict delays in communication networks, like queu‑
is the basic setting of Graph Neural Networks (GNNs). ing theory [6], network calculus [7] and simulation‑based
GNNs use the so‑called ”Message Passing” algorithm [2] approaches [5].
and can express the notion of nodes and edges. However, Boutaba et al. [8] provide a general overview on the appli‑
for communication networks it is also important to con‑ cation of machine learning to communication technolo‑
sider paths (and network lows) along several consecu‑ gies and network measurements. The approaches in par‑
tive links. RouteNet [3, 4] is an implementation of this ticular differ in whether the data used for learning comes
idea that allows expressing paths. The RouteNet archi‑ from network simulators (e.g. from OMNeT++ as in our
tecture consists mainly of two gated recurrent neural net‑ case) or from actual measurements. In addition, they can
works that are responsible for calculating path and link be divided into supervised, unsupervised and reinforce‑
properties, respectively. ment learning. The approach considered in this paper is
The RouteNet architecture can be used to predict per‑ an example of supervised learning.
low performance metrics such as average delay and jit‑ Mestres et al. [9] investigate modeling and prediction
ter. This can be useful for assessing networks with re‑ of delays in communication networks with feed‑forward
spect to different loads without needing to test them in neural networks. They predict the latency based on the
reality. Hence, it is possible to determine if a network traf ic con iguration. In contrast to the RouteNet archi‑
can handle a certain load with respect to a performance tecture, a neural network has to be trained for each spe‑
metric such as average delay. An alternative to such a ci ic communication network. Graph neural networks
prediction with neural networks is a simulation, using and message passing were irst introduced by Scarselli et

© International Telecommunication Union, 2021 1