RouteNet: Leveraging Graph Neural Networks for
- F% J9 @. b4 LNetwork Modeling and Optimization in SDN , E! c+ \: c. R1 P- B, Z5 D2 Y* i
! b& w* m2 Y. K! h7 r6 {: S& A
% d- X6 G6 h% ]0 nNetwork modeling is a key enabler to achieve
$ M4 \* B- v- M8 `* L. Zeffificient network operation in future self-driving Software) L$ {2 C& h4 {9 P! J$ t
Defifined Networks. However, we still lack functional network
0 S) d( {* H; [2 X# u2 Tmodels able to produce accurate predictions of Key Performance
( V& C; y7 L1 ^* K- b lIndicators (KPI) such as delay, jitter or loss at limited cost.' `, K7 d( g. q
In this paper we propose RouteNet, a novel network model based6 I* ^& V9 S. \+ M4 \
on Graph Neural Network (GNN) that is able to understand) `( j. o3 @4 Y! T. {- }
the complex relationship between topology, routing, and input
0 v' Z9 i; H5 i) F2 q Qtraffific to produce accurate estimates of the per-source/destination: c- q) [2 i* `) ?8 t, A) H
per-packet delay distribution and loss. RouteNet leverages the& J% o2 a! q' q) n8 N9 M
ability of GNNs to learn and model graph-structured information
# {% U9 N- v9 ^$ d6 Wand as a result, our model is able to generalize over arbitrary7 b/ x( Y- X5 c" y
topologies, routing schemes and traffific intensity. In our eval/ b/ E. o4 U8 ^: L7 S
uation, we show that RouteNet is able to predict accurately0 X1 d7 T1 B& t! P- O
the delay distribution (mean delay and jitter) and loss even in+ [. n" ?" @* N1 `, j
topologies, routing and traffific unseen in the training (worst case+ x( F( [% G* Q7 G& u* |# ^
MRE = 15.4%). Also, we present several use cases where we* }! n3 Z* p+ c6 q' |: o6 K; f
leverage the KPI predictions of our GNN model to achieve
9 w0 ~$ H2 ^5 Y- Ieffificient routing optimization and network planning.
* j7 s4 Y8 T9 \6 s$ {: F. z5 Q6 |" v/ I. o O( X# m
& e7 l; [8 \/ {% ` \ |