RouteNet: Leveraging Graph Neural Networks for
2 j5 ]: x' a4 J& h. h H( V. }6 BNetwork Modeling and Optimization in SDN . l2 M7 o8 m1 o1 E- g/ A
) I& D0 C& F% P! {" s% ]
0 p) l9 n# q" V. l9 U# P ?Network modeling is a key enabler to achieve
; J( z }3 j$ v- Q1 z6 Neffificient network operation in future self-driving Software
7 c* {0 Y) o4 J3 xDefifined Networks. However, we still lack functional network
2 n" W7 c; m; r x6 J4 @models able to produce accurate predictions of Key Performance* z- e; x! s$ c9 U
Indicators (KPI) such as delay, jitter or loss at limited cost.
; s$ Q. m' [9 [4 E/ K5 kIn this paper we propose RouteNet, a novel network model based
; h( O+ M2 V4 C* ton Graph Neural Network (GNN) that is able to understand7 [) c# L( J( M. Z
the complex relationship between topology, routing, and input
$ y4 f, i& P9 ~& W) t/ vtraffific to produce accurate estimates of the per-source/destination
8 Q1 Q% U5 Z7 W) h4 P. F4 L* Tper-packet delay distribution and loss. RouteNet leverages the7 x- ~8 I1 b& t3 }) w- |
ability of GNNs to learn and model graph-structured information6 }+ C( T4 A) K4 e% X+ o( a
and as a result, our model is able to generalize over arbitrary
- n. a( o- u7 P2 j7 Utopologies, routing schemes and traffific intensity. In our eval
2 P9 a [: t4 M1 Cuation, we show that RouteNet is able to predict accurately/ ?) z0 w' n# v" F Y; ?
the delay distribution (mean delay and jitter) and loss even in/ Z& l0 f' g% t' L0 W: J; d0 K
topologies, routing and traffific unseen in the training (worst case
5 q4 a$ e$ g) H( ?0 pMRE = 15.4%). Also, we present several use cases where we
% u# F+ P9 n: }5 s% S& `# `+ aleverage the KPI predictions of our GNN model to achieve
- P) L1 y; D! L% Yeffificient routing optimization and network planning.
w/ N# g: S$ Q. }8 s
7 o7 O% R) C% I: Y1 [; D# v; J' v$ l
|