Modeling Fil e ing Penal ies in ROADM-based Ne wo ks
wi h Machine Lea ning o QoT Es ima ion
Ankush Mahajan(1)*, Kos as Ch is odoulopoulos(3), Rica do Ma inez(1), Sal a o e Spada o(2), Raul Munoz(1)
(1) Cen e Tecnològic de Telecomunicacions de Ca alunya (CTTC/CERCA), (2) Poly echnic Uni e si y o Ca alonia (UPC), Ba celona, Spain,
(3)Nokia Bell Labs, S u ga , Ge many
*ankush.mahajan@c c.ca
Abs ac : Moni o ing 3dB bandwid h and o he spec um ela ed pa ame e s a ROADMs p o ides
in o ma ion abou quali y o hei il e s. We p opose a machine-lea ning model o es ima e end- o-
end il e ing penal y o mo e accu a e QoT es ima ion o u u e connec ions. © 2020 The Au ho (s)
OCIS codes: 060.4256 Ne wo ks, ne wo k op imiza ion; 060.4510 Op ical communica ions
1. In oduc ion
Recon igu able Op ical Add/D op Mul iplexe s (ROADMs) a e he key swi ching elemen s o deployed co e and me o
op ical ne wo ks [1]. Se e al implemen a ions o ROADMs a e possible using op ical de ices including
MUX/DEMUX, op ical spli e s/combine s, wa eleng h blocke s and wa eleng h selec i e swi ches (WSSs) [2]. The
use o WSS p o ides he ad an ages o colo less, and/o di ec ionless and/o con en ionless node ope a ion and highe
deg ee design, making hem indus y’s choice o cu en gene a ion ROADMs [1, 2].
Gene ally, a signal ha a e ses a ROADM node su e s om il e ing penal y due o he in ol ed WSS(s)
esul ing in signal quali y o ansmission (QoT) deg ada ion. O e longe pa hs he cascade o ROADM il e s
in oduces s onge il e ing, which na ows he signal’s ansmission bandwid h (BW) [3]. Du ing he ligh pa h
p o isioning , such op ical il e ing penal y is co e ed inside ma gins o he QoT es ima ion ool (o Q ool). Resea ch
wo ks [3, 4] con i med ha ROADM node penal y (in OSNR) inc eases exponen ially wi h he numbe o nodes and
depends on he modula ion o ma and he g id spacing. An analy ical model based on a highe o de SNR-OSNR
ela ion o cap u e cascaded il e ing e ec s is p esen ed in [5]. These p io wo ks ocus on cha ac e iza ion o he
cascade (ei he by looping single il e , o by eplacing cascaded il e s wi h a unable BW il e ) and do no ocus on
he iden i ica ion o he quali y o he indi idual il e s. Al hough hese wo ks consen ha il e ing penal y is non-
linea o he numbe o il e s, s ill he e a e unce ain ies in hese penal ies. The misalignmen o he il e s o he g id,
de ia ions in il e s shape and 3dB BW a e qui e common in deployed ne wo ks and a e co e ed in he a o emen ioned
Q ool ma gins. Such issues a e expec ed o exace ba e in disagg ega ed op ical ne wo ks whe e ROADM/ il e s and
Tx. lase s could come om mul iple endo s wi h di e se cha ac e is ics [5, 6]. In disagg ega ed (a node le el)
ne wo ks ma gins o 3.5-5dB (co e) and 3-3.5dB (me o) would be equi ed [6] mainly due o unce ain ies/ a iabili y
o he mul i- endo componen s; he il e ing penal y o ROADM nodes play a signi ican pa in hose inc eased
ma gins. In a ne wo k (ei he single o mul i- endo ), connec ions a e se di e en ROADMs and expe ience di e en
deg ees o il e ing penal y. Using cheap op ical channel moni o s (OCM) [7] we can unde s and he il e s alignmen
o he g id and hei shape. Such unde s anding could be used o co ec he Tx/cascade alignmen and imp o e he
QoT o exis ing connec ions [8] o imp o e QoT es ima ion and educe ma gin o u u e connec ion es ablishmen s.
In ligh o he abo e he ein, we i s ly examine il e s beha io unde unce ain ies and co esponding penal ies.
We hen p opose a Machine Lea ning (ML) eg ession model based on link o mula ion ha le e ages moni o ing da a
o es ablished connec ions o accu a ely es ima e end- o-end il e ing penal y o new connec ion eques s.
2. Me hodology and P oposed Solu ion
ROADMs consis s o ampli ie s and il e s o boos and ou e he signal. Now, conside a channel ha ing cen al
wa eleng h 𝝀 ha is ou ed h ough mul iple ROADMs o e a pa h p be o e inally de ec ed a he ecei e . This can
be iewed as a line ansmission sys em wi h il e s along he pa h (1 o 2 pe ROADM o B oadcas and Selec -B&S,
o Swi ch/Rou e and Selec -S&S a chi ec u e). We deno e he spec um o an indi idual il e in he pa h as 𝑭𝒊(𝝀), i∈
𝒑 and by pi he pa o he pa h be o e i h il e . We also deno e by 𝑪𝒑𝒊(𝝀) he o e all il e spec um be o e i h il e due
o he cascade o p e ious il e s o e pi and by 𝑪𝒑(𝝀) he o e all il e spec um a end o he pa h (Rx.), is gi en by:
𝐶𝒑𝒊= ∏𝐹𝑘(𝜆)
𝒌∈𝒑𝒊, 𝐶𝑝(𝜆) = ∏𝐹𝑘(𝜆)
𝒌∈𝑝 (1)
Assuming ha i' is he nex il e a e i on he pa h, by moni o ing 𝐶𝑝𝑖(𝜆) and 𝐶𝑝𝑖′(𝜆) we can calcula e 𝐹𝑖(𝜆) which
gi es us aluable in o ma ion abou key p ope ies ( il e shape and alignmen o he g id) o he i h il e a wa eleng h
𝜆. Typically, such p ope ies hold o all 𝜆s o he same il e . F om moni o ed 𝐶𝑝𝑖(𝜆), we ex ac a se o ea u es j ha
e lec s he p ope ies o cascade be o e i, deno ed by 𝐶𝑖
𝑗(𝜆), such as 3dB BW, cascaded il e cen al equency, signal
dis ibu ion pa ame e s (1s /2nd o de s a is ical momen s), o 𝐶𝑖
𝑗(𝜆) = {𝐶𝑖
3𝑑𝐵(𝜆), 𝐶𝑖
𝑓𝑐(𝜆), …, 𝐶𝑖
𝑠𝑦𝑚.(𝜆)}. Figu e 1(a) shows
Fig. 1: (a) OSNR penal y & 3dB BW, (b) spec um o inc easing numbe o cascaded WSSs (iden ical 𝐹𝑖(𝜆) o all WSSs), (c) spec um o 3dB
BW unce ain y 𝛥𝑖
3𝑑𝐵 =±10% (non-iden ical 𝐹𝑖(𝜆)), esul ing in ~2.1GHz o unce ain y a 5 h cascaded WSS, compa e o Fig. 1(b) ( ed egion)
he non-linea /exponen ial deg ada ion o one such ea u e, i.e. 3dB BW, 𝐶𝑖
3𝑑𝐵(𝜆), in a cascade o iden ical il e s (3dB
BW=37.5 GHz, 2nd o de Gaussian shape) ob ained wi h simula ions in VPI. I also shows he OSNR (in dB) penal y
o h ee modula ion o ma s (@ 32Gbaud, α=0.1). Focusing on spec al esponse, Fig. 1(b) shows he spec al shape
o he signal 𝐶𝑝𝑖(𝜆), which deg ades as he numbe o cascade inc eases, e en o iden ical il e s. We obse ed a 3dB
BW deg ada ion o 𝐶𝑖
3𝑑𝐵(𝜆)=6.02GHz a e 5 iden ical il e s. Howe e , in eal ne wo ks sligh a ia ions a e ypical
wi hin he spec al esponses e en o iden ical il e s, while such a ia ions would exace ba e in disagg ega ed
scena ios i il e s/ROADMs come om di e en endo s. Such a ia ions esul in unce ain ies in he ea u es 𝐶𝑖
𝑗(𝜆),
which we deno e as 𝛥𝑖
𝑗(𝜆) = {𝛥𝑖
3𝑑𝐵(𝜆), 𝛥𝑖
𝑓𝑐(𝜆), . . , 𝛥𝑖
𝑠𝑦𝑚.(𝜆)}. Fig. 1(c), shows he esul ed 𝐶𝑝(𝜆) o a 3dB BW unce ain y
o 𝛥𝑖
3𝑑𝐵=±10% pe il e a e a cascade o 5 il e s, simula ed in VPI. We obse ed ~2.1GHz unce ain y in he 3dB
BW, 𝐶𝑝
3𝑑𝐵(𝜆), a he end o he cascade/pa h, which con ibu es o inaccu a e il e penal y es ima ion.
Fig. 2: (a) dis ibu ed OCM loca ions o a sample ne wo k (4 nodes, wi h es ablished connec ion om A o D) wi h swi ch & selec , S&S
ROADM a chi ec u e (in inse ), (b) end o end (link o mula ion app oach based) ML model along wi h ea u e ma ix, X and a ge ec o , e
Focusing on il e penal y modeling, a s anda d Q ool (deno ed by Qs) calcula es he ea u es (e.g. 3dB BW) and il e
penal y along a pa h assuming iden ical il e s (Fig. 1(a)) and uses a high ma gin on op o accoun o penal ies om
inaccu acies ( il e alignmen and shape). By exploi ing moni o ing in o ma ion, we can ex end Qs o unde s and he
ac ual ne wo k s a e and beha io o deployed il e s, educe inaccu acies and lowe he ma gin o new connec ions
[9, 10]. Today cheap OCMs [7] can be ins alled a ROADM nodes o moni o 𝐶𝑝𝑖(𝜆) on any channel. Each ROADM
includes WSS/ il e s depending on i s deg ee and a chi ec u e. Depending on add/d op o c ossing di ec ion, di e en
il e s a e encoun e ed in he ne wo k, hence in he ollowing we accoun he il e s i on a pe link bases. We assume
an op ical ne wo k wi h es ablished connec ions and hei a ibu es deno ed by P. We use OCMs o moni o 𝐶𝑖(𝜆𝑃),
and ex ac 𝐶𝑖
𝑗(𝜆𝑃) be o e each ROADM node along he pa hs o es ablished connec ions (P). So, he ea u es 𝐶𝑖
𝑗(𝜆𝑃
supp ess o simpli y he no a ion) se e as he g ound u h and a e s o ed in Q ool da abase. Typically, a s anda d
Q ool would s a om he Tx. pa ame e s and i e a i ely calcula e he cascaded ea u es o a speci ic il e Fi wi h no
unce ain y, 𝛥𝑖
𝑗=0, down he pa h un il he ecei e and add a ma gin o no accoun ing o unce ain ies. So he
s anda d Q ool, Qs, includes a unc ion ha akes he ea u es be o e ROADM i, 𝑐𝑖
𝑗, and calcula es he expec ed ea u es
a e ROADM i, Qs(𝑐𝑖
𝑗, 𝐹𝑖) assuming no unce ain y 𝛥𝑖
𝑗=0 (Fig. 1(a) shows such a Qs unc ion). Since we ha e OCM
in o ma ion 𝐶𝑖
𝑗, we can co ec Qs and educe he ma gin as ollows. We deno e he expec ed ea u e as 𝑐𝑖′
𝑗=Qs(𝐶𝑖
𝑗, 𝐹𝑖)
and he moni o ed-expec ed e o as 𝑒𝑖′
𝑗= 𝐶𝑖
𝑗 - 𝑐𝑖′
𝑗 , which is due o unknown unce ain ies 𝛥𝑖
𝑗. Then we ex ac a pe
link ea u es ma ix, 𝑋𝑖
𝑗 = (𝐶𝑖
𝑗, 𝑐𝑖
𝑗). Ou goal is o iden i y a pe link dependen e o unc ion, Θ(𝑋𝑖
𝑗) ≈ 𝑒𝑖
𝑗, which
maps he ea u es ma ix 𝑋𝑖
𝑗 o he e o 𝑒𝑖
𝑗. We ely on ML o aining and i ing o X on e and inding Θ. Fig. 2(b)
shows he ea u es ma ix X u ilizing OCM da a 𝐶𝑖
𝑗, Qs expec ed ea u es 𝑐𝑖′
𝑗 and e o e o he oy ne wo k o Fig.2(a).
Assuming a new connec ion eques p
∉
P using wa eleng h l, we s a wi h i s ansmission spec um pa ame e s
𝑐𝑝0
𝑗(𝑙), use Qs o ob ain he expec ed ea u e se 𝑐𝑝1
𝑗(𝑙) a e he ing ess node, ex ac 𝑋𝑝0
𝑗(𝑙) co ec ha by calcula ing
(a.)
18
20
22
24
26
28
30
32
34
0
2
4
6
8
10
12
14
16
020 40
E ec i e 3dB BW (GHz)
OSNR penal y (dB)
No. o cascaded WSS
DP-QPSK
DP-8QAM
DP-16QAM
3dB BW
36
34
32
30
28
26
24
22
40
16
14
12
10
8
6
4
2
00
DP-QPSK
DP-8QAM
DP-16QAM
3dB BW
(a.)
100 300 500 700 900
6.02GHz
Tx. signal 1 WSS
3 WSS 5 WSS
-15
-20
-25
-30
-35
-40
(b.)
OSNR penal y (dB)
No. o cascaded WSS
20
Powe (dB)
F equency ela i e o 193.1THz (GHz)
-36 -18 018 36
150 350 550 750 950
2.1GHz
F equency ela i e o 193.1THz (GHz)
Tx. signal
5 WSS wi h Δi, 0%
5 WSS wi h Δi, 10%
-18 018 36
-36
(c.)
-15
-20
-25
-30
-35
-40
Powe (dB)
20
E ec i e 3dB BW (GHz)
W
S
S
W
S
S
ROADM
OCM: Op ical Channel Moni o s
ROADM
sample connec ion:
A→D
ROADM
ROADM
ROADM
OCM
A
B
C
D
ea u e ma ix, X o il e ing unce ain ies
-----
-
-----
-
Modula ion Fo ma , M.F #1
-----
-
-----
-
M.F #2 M.F #M
P, connec ions/ligh pa hs
(node based pe link ea u es)
links links
M
a
p
p
i
n
g
ML
Model
Θ=
(b.)
𝑪
𝑪
𝑪𝑪
𝒊
TXBW_A F
u
n
c
i
o
n
moni o ed expec ed C
C
------
------
------
------
𝑪
𝑪𝑪
𝑪𝒊
link IDs AB BC CD links links
calcula ed
e o /M.F
a ge
ec o , 𝒊
pe link
𝑪
𝒊′
------
(a.) (SVM)
𝑪
𝑪
𝒊′
𝑒𝑝0
𝑗= Θ(𝑋𝑝0
𝑗(l)), and epea ha link by link down he pa h un il des ina ion. The es ima ion e o will be iden i ied once
we es ablish he connec ion, moni o 𝐶𝑝𝑖
𝑗(𝑙) a a ailable OCMs and compa e i o es ima ions by he abo e algo i hm.
3. Resul s & Discussion
To quan i y he bene i s o de eloped QoT es ima o wi h mo e accu acy and educed ma gins, we conside ed DT
opology wi h 12 nodes and 40 bidi ec ional links wi h leng hs om 48 o 458 km as (inse Fig. 3(a)). The span was
assumed o be s anda d single-mode ibe and span leng h equal o 80km. Each demand is ca ied by one wa eleng h
and modula ed a 32Gbaud wi h {QPSK, 8-QAM, 16-QAM} modula ion o ma s leading o {100, 150, 200} Gbps o
da a a e. The equency slo size was assumed o be 12.5GHz ( ixed) and we alloca ed 3 spec um slo s o 12.5GHz.
Fig. 3: (a) end o end e ec i e 3dB BW e o and educed ML es ima ed e o , (b) e o s (± e) in OSNR (dB): e e ence (black), ML max
o e es ima ion wi h OCM da a (g een), (c) new ma gin & hei educ ion wi h di e en Δi in ensi ies
We assumed a ne wo k wi h OCMs ins alled a each node (Fig. 2(a)) and gene a ed moni o ing da a (g ound u h), by
andomly applying small Δi ( esul ed in end o end ±1.5GHz 3dB a ia ion) e lec ing Tx. & il e s-g id misma ch, and
small a ia ions in il e s shape. The co esponding OSNR penal ies a e also dis ibu ed in ± e sides depending upon
Δi. +/- e penal ies esul in uppe /lowe bound o design ma gins and we call hem as, “high/low ma gin o e o s”.
We assumed a s able ne wo k s a e, whe e a se o connec ions is es ablished and he aim is o p o ision a new se o
new connec ions. To do so, we di ided he connec ions in o wo se s o aining and es ing, assumed o be he
es ablished and he new connec ions, espec i ely. Then om he aining da ase , we calcula ed e o s, e based on he
expec ed and moni o ed 3dB BW, cen al wa eleng h, symme y, link IDs, ou e e c. in o ma ion. We also gene a ed,
he pe link ea u e ma ix, X. We used suppo ec o machine, SVM, i ing echnique wi h gaussian ke nel unc ion,
o ain ou ML model, and we achie ed leas max. MSE o ~0.02 dB on p edic ed OSNR a a maximum load o 400
connec ions (200 imes a e age @ 400 connec ions). Fig. 3(a) shows he calcula ed e o in 3dB BW a a load o 400
connec ions and also he e o educ ion in BW ( om ±1.5GHz → ~0.18GHz) wi h ained SVM. Fig. 3(b) e lec s
hese accu a e (pe link) es ima ion o 3dB BW in end- o-end accu a e es ima ion o il e ing/OSNR penal y (g een
lines). Fig. 3(b) shows ha a 90%/10% ain/ es spli , maximum e o educ ion/accu acy imp o emen was ~0.67dB
o high e o and ~0.68dB o low e o s, espec i ely. These a e he new educed high and low ma gins o he ac ual
and unknown il e ing unce ain ies. Fo high/low ma gin, we ound an o e all educ ion o 80.4/83.4% a a load o
400 connec ions. We also a ied Δi (mul iplied by a ac o o 1/3 o 3) and es ima ed high and low ma gins/e o s a a
ixed load o 400 connec ions. In Fig. 3(c), he high Δi scena io (> 1dB, igh o ed dashed line) e lec s ROADMs
nodes wi h highe unce ain y, which a e expec ed in disagg ega ed/mul i- endo ne wo ks. As expec ed, highe
e e ence ma gins a e equi ed he e, and ou accu a e modeling esul s in mo e p onounced sa ings ha each >85%
and >1.5 dB on bo h high and low ma gins.
4. Conclusion
We p oposed ML model o es ima e end- o-end penal y gene a ed a ROADM nodes due o il e spec al unce ain ies
& hei cascaded e ec s. Ha nessing moni o ed da a and le e aging ML echniques, we es ima ed QoT accu a ely o
new connec ions wi h max. o ~0.68dB o OSNR accu acy and >80% educ ion in ela ed ma gin.
Acknowledgemen s: Au ho s would like o hank Ka s en Schuh and Camille Delezoide o Nokia Bell Labs o echnical discussions
on il e modelling. This wo k is a pa o H2020-MSCA, ONFIRE p ojec suppo ed by EU, g an ag eemen No. 765275.
5. Re e ences
[1] M. File , e al., “N-deg ee ROADM A chi ec u e Compa ison: B&S s. R&S in 120 Gb/s DP-QPSK T ansmission Sys ems,” OFC, 2014
[2] B. Cloue , e al., “Ne wo king Aspec s o Nex -Gene a ion Elas ic Op ical In e aces,” JOCN, 2016
[3] J. M. Fab ega, e al., “On he il e na owing issues in elas ic op ical ne wo ks,” JOCN, 2016
[4] T. Rahman, e al., “On he Mi iga ion o Op ical Fil e ing Penal ies O igina ing om ROADM Cascade,” PTL, 2014
[5] C. Delezoide, e al., “Weigh ed Fil e Penal y P edic ion o QoT Es ima ion,” OFC, 2018
[6] M.P. Belange , e al., “Ma gin equi emen o disagg. he DWDM anspo sys. and i s consequence on applica ion economics,” OFC, 2018
[7] h ps://www. inisa .com/ oadms-wa eleng h-managemen / ocm01 xc1mn
[8] C. Delezoide, e al., “Au oma ed Alignmen Be ween Channel and Fil e Cascade,” OFC, 2019
[9] K. Ch is odoulopoulos, e al., “Towa d e icien , eliable, and au onomous op ical ne wo ks: he ORCHESTRA solu ion [In i ed],” JOCN, 2019
[10] A. Mahajan, e al., “Machine Lea ning Assis ed EDFA Gain Ripple Modelling o Accu a e QoT Es ima ion,” ECOC, 2019
1.5
-0.75 0.75
-1.5 0
BW e .
educed o
<0.2 GHz wi h
mean, µ ≈ 0.02
(a.)
0
0.6
0.4
0.2
p obabili y
p ed. BW e .
ac . BW e .
cascaded 3dB BW e o (GHz)
-1.5
-1
-0.5
0
0.5
1
0.5 0.6 0.7 0.8 0.9
OSNR (dB) o e es .
%age o aining da ase @ 400 connec ions
e . low e o e . high e o
es . low e o - OCM es . high e o - OCM
-2
-1
0
1
2
0.2 0.7 1.2 1.7 2.2 2.7 3.2
OSNR (dB) o e es .
Δidi iso @ 400 connec ions
e . low e o e . high e o
es . low e o - OCM es . high e o - OCM
~0.67dB
~0.68dB
~0.46dB
~0.42dB
Ma gin Reduc ion, M.R > 87%
M.R > 85%
low Δi/
wi hin ~1dB
ma gin
high Δi/ high ma gins
owa ds disagg ega ion