RouteNet: Leveraging Graph Neural Networks for
) D- W# x2 G: ?% o! cNetwork Modeling and Optimization in SDN ' u2 z( d0 ~2 {, I9 S& a4 ~
, I, u7 V+ w# T8 y) H/ P' j/ K
0 _9 D% X% j- F# a+ hNetwork modeling is a key enabler to achieve. u& ]5 Z' j5 D9 C
effificient network operation in future self-driving Software
9 \. \5 N4 I1 Y! e& f) e- G& E" jDefifined Networks. However, we still lack functional network
! G/ V! L A3 D( B& J9 Ymodels able to produce accurate predictions of Key Performance2 E _3 u+ a! K6 f2 g1 ~
Indicators (KPI) such as delay, jitter or loss at limited cost.
% L/ k; T$ Z% p" i7 Y0 V3 gIn this paper we propose RouteNet, a novel network model based1 G7 o1 s# G5 A2 L) X3 g3 N6 J" Q
on Graph Neural Network (GNN) that is able to understand
' b: B! Z0 Z+ @( A, @/ ]the complex relationship between topology, routing, and input
* r' z8 K2 w, f$ E$ c) b$ otraffific to produce accurate estimates of the per-source/destination! f( r6 I: H: u: X
per-packet delay distribution and loss. RouteNet leverages the
2 z' p2 m1 h! L/ }7 K+ H( A5 b- m% qability of GNNs to learn and model graph-structured information J% y9 H" B5 E4 ^. A& m3 Z
and as a result, our model is able to generalize over arbitrary
: M& E9 }$ V+ P) F$ H; {8 Wtopologies, routing schemes and traffific intensity. In our eval: F. i" e3 X; b2 v* z+ a/ l( w
uation, we show that RouteNet is able to predict accurately
, ~* M% E, h& H. y9 Xthe delay distribution (mean delay and jitter) and loss even in# h1 r) V2 n) j6 C( G* s. I4 B
topologies, routing and traffific unseen in the training (worst case' W6 s; }1 D, K7 Z
MRE = 15.4%). Also, we present several use cases where we- F Y' q [: p& `% H- B* M
leverage the KPI predictions of our GNN model to achieve
) X- U# U+ a5 P/ E5 |9 {effificient routing optimization and network planning.
7 T( _" ? q% I, c' |3 P. {0 |1 u! f( ^( \
1 G8 t! O2 w% i& o7 v |