Page 18 - 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

In communication networks, scheduling policies describe
Network topology in which order packets are transmitted. A simple and
Traffic information straightforward algorithm is FIFO, where the packets are
Neural network
GNN model Average per-path delay
Routing configuration transmitted in the order in which they are received [17].
Scheduling configuration The original RouteNet was developed for networks that
use only a single scheduling policy. However, this im‑
plementation does not work well with other scheduling
Fig. 1 – Problem setting for the challenge algorithms and networks with heterogeneous schedul‑

al. [1] and Gilmer et al. [10]. These concepts were applied ing can have large a impact on the behavior of networks
to the domain of communication technologies by Geyer and thus on the delays. For this challenge speci ically,
and Carle [11], where the authors use a GNN for automatic three different scheduling policies, namely Weighted‑
network protocol design. Fair‑Queuing (WFQ), Strict Priority (SP), and De icit
A different approach to deal with heterogeneous schedul‑ Round Robin (DRR) are being used. For WFQ and DRR,
ing policies was recently proposed by Ferriol et al. [12]. there are three ToS classes. The networks are in general
They provide a GNN architecture with states for links, not homogeneous with respect to scheduling policies, in
paths and queues. This model reaches a mean relative fact there are data sets where all policies are present in a
error of 3.88% in the German Backbone Network (GBN) single communication network.
topology (which was not used for the training process).
3.1 RouteNet
3. SETTING RouteNet uses Graph Neural Networks and the so‑called
The solution proposed in this paper was developed within message passing [2] for predicting average per‑path de‑
the GNN Challenge [13] provided by the ”Barcelona Neu‑ lays in communication networks. There are two impor‑
ral Networking Center” from the Universitat Politècnica tant elements in this architecture, links and paths; where
de Catalunya. This challenge was organized as part of the each path consists of at least one link. Note that we use
”ITU Arti icial Intelligence/Machine Learning in 5G Chal‑ the term ”capacity” to refer to the tight upper bound on
lenge”. the transmission rate of a link and the term ”data rate” to
The goal of the GNN challenge was to predict the aver‑ refer to the desired transmission rate of a traf ic low on a
age per low delay in a communication network. In other network path. Each link is associated with speci ic infor‑
words, it is of interest to estimate the average time it takes mation, such as link capacity. We will refer to this simply
for a packet to travel from its source node to its desti‑ as link state information, which is represented as a vector.
The same holds true for paths, which we will call analo‑
nation node. Additionally, the network topology may be
gously path state information. The RouteNet version pro‑
different from that used in the training data. Thus, the
vided for the challenge [18] uses at initialization only link
neural network should not be adapted to only one topol‑
ogy but work for general topologies. For this, three differ‑ capacity (bits/time unit) for the link state information ℎ
and the average data rate (bits/time unit) of a single low
ent network topologies have been provided. For training,
the NSFNET topology with 14 nodes [14] and GEANT2 for the path state information ℎ . The data rate of the low

topology with 24 nodes [15] were used. For validation, can be encoded as part of the path state information as in
GBN [16] with 17 nodes was used. The data set that was RouteNet it is assumed that there is at most one low per
used for the evaluation of the challenge consisted of 19 path. Thedimensionof linkand pathinformationareboth
nodes. Other information has not been published by the set to 32 in this RouteNet version, that is at the beginning
challenge organizers. Thus, the proposed solution for this only one component of the link and path state contains
challenge must work even for such unknown communica‑ meaningful information, all other components are illed
tion networks where no details are known beforehand. with zeros. Therefore, those two vectors can be written
The data set consisted of different node and link informa‑ as
tion, as well as different settings used in the OMNeT++ ℎ = [ , 0, … , 0] ∈ ℝ 32 and
′

simulator. A crucial simpli ication for all data sets is that ℎ = [ , 0, … , 0] ∈ ℝ ,
′
32
there exists only one low for each path. And for each low,
a Type of Service (ToS) is randomly assigned. That means where denotes the link capacity and the average de‑
all packets of a path have the same ToS. This is an impor‑ sired data rate on that path.
tant property of the data set and we will utilize it in Sec‑ RouteNet utilizes two recurrent neural networks and

tion 4.3. The number of generated packets per time unit . The neural network calculates the new path state

follow a Poisson distribution, and the inter‑packet arrival information based on the previous path information and
times follow an exponential distribution. A two‑valued link information. The result of this neural network is then
distribution is used to model packet size. The maximum used for to calculate new link state information with

bit rate is chosen randomly between 400 and 2000 bits the previous link state information. See Algorithm 1 for
per time unit. For more details on the simulated data, we the pseudo‑code. As is a recurrent neural network,

refer to the challenge documentation [13]. it returns the hidden path state after each link of a path.

2 © International Telecommunication Union, 2021