1

Cognitive Power Management for Optical

Superchannel Systems
Valery N. Rozental and Darli A. A. Mello

Abstract—We presented a scheme for cognitive power man- A challenge is to map the aggregate traffic for the line-side
agement (CPM) in software-defined, superchannel optical transmission in the most efficient way given the available
transponders. The scheme reduces the length of electronic
bandwidth resources. Here, a key component is a software-
chromatic dispersion compensation filters according to the
network traffic fluctuations. Transmission rate reduction is defined optical transponder that supports variable trans-
achieved by symbol retransmission, while maintaining inter- mission rates and transmission parameter adjustment (e.g.
nal parameters of the electronic devices unchanged, to avoid symbol rate, modulation format, DSP algorithms) according
increased complexity. We investigate single and multiple to the optical channel conditions: optical signal-to-noise ratio
carrier transmission schemes in terms of Pareto-efficiency
regarding transmission rate, reach and power consumption. (OSNR), available optical bandwidth, chromatic and polar-
We show that, for the same transmission rate, polarization ization mode dispersions. Aside from the obvious efficiency-
multiplexed and higher modulation order schemes are more related benefits, single transponder type-based networks
power efficient. We further investigate the optimal number reduce costs associated with certification, testing, training
of active subcarriers in a CPM-capable superchannel-based
optical line interface, showing that Pareto-efficient schemes
and spare part inventory [4].
keep active over half the subcarriers. We also assess the Commercial product development for 100G optical sys-
optimization of the number of subcarriers in a superchannel tems raised an important issue of power consumption in
transponder when each new subcarrier results in some addi- optical transponders, leading to such solutions as single
tional complexity – even when non-active. We show that the chip ADC/DSP, use of advanced CMOS processes and optical
influence of the additional complexity increases with OSNR
and propose to set the number of subcarriers after OSNR integration. Higher per channel transmission rates like 400
estimation is performed for the specific link. We further Gbps and 1 Tbps will rely, at least in the near future, on op-
propose to implement subcarrier sleep-mode by transmitting tical parallelism that allows to overcome electronic devices’
at a minimum rate using QPSK at 3.5 Gbaud, so that – while bandwidth limitations. This parallelism further stresses the
maintaining a significant complexity reduction – there is
no discontinuity in clock, state of polarization and phase
importance of power consumption optimization, because sys-
tracking, and the subcarrier can be smoothly reactivated. tem complexity is multiplied in optical superchannels [5].
Finally, we conduct a case study to demonstrate power Therefore, we view power efficiency of different transponder
savings offered by CPM in a typical terabit transmission architectures as a figure of merit.
scenario.
Consider a software-defined, superchannel-based optical
Index Terms—Optical transponders; bandwidth virtualiza- transponder that supports different transmission rates and
tion; cognitive power management; chromatic dispersion modulation formats, where operator can also define a num-
compensation; Pareto-efficient transmission.
ber of active subcarriers. Various combinations of these pa-
rameters form a set of the supported transmission schemes.
The suitable scheme is chosen during link configuration
I. I NTRODUCTION
according to the required transmission rate and the optical

N owadays, optical communication systems undergo a

major transmission capacity upgrade, achieving data
rates beyond 100 Gb/s per wavelength due to the employ-
signal-to-noise ratio (OSNR), to guarantee a certain bit-
error rate (BER). If for specific channel conditions more than
one scheme attend the required rate and BER criteria, the
ment of new technologies, namely, optical coherent detection, choice should be based, according to our figure of merit, on
advanced modulation formats and polarization multiplexing. minimizing the power consumption.
Along with capacity, network efficiency is gaining attention, Once the transmission scheme is chosen, the transpon-
addressing system cost, transparent reach, quality of service, der transmits at a constant nominal rate, to accommodate
power consumption, failure tolerance, optimized bandwidth the highest traffic load. However, internet traffic presents
allocation and efficient routing. This multi-faceted task in- considerable fluctuations: Figs. 1(a-b), illustrates daily and
volves further developments in the optical layer technologies, weekly profiles of the aggregate internet traffic for all ex-
protocols and multi-level integration, management and con- change points between autonomous systems networks in
trol implementation [1]. Some solutions require a paradigm Brazil [6]. There exists a substantial fluctuation between the
shift: for example, ITU-T recommendation G.694.1 from daytime and the nighttime traffics. Weekly traffic also ex-
February 2012 [2] introduces a high granularity, 12.5 GHz hibits fluctuations, having a reduction of almost 40% during
DWDM grid, and a multi-slot flexible grid, abandoning the weekends. Although actual values may vary from one link to
traditional 50/100 GHz channel spacing. Efficient bandwidth another, the general pattern of these graphs should remain
use is achieved through capacity virtualization, where the similar. These fluctuations, if properly explored, may result
service layer is decoupled from the optical transmission in system optimization. Thus, by lowering transmission rates
layer. Here, instead of attributing a specific rate and wave- and choosing dynamically the transmission scheme we can
length for each service while dealing with the variety of reduce the length of the digital filters used for CD compensa-
associated protocols, the WDM line interface processes exclu- tion [7], reducing computational complexity and power con-
sively the aggregate traffic, allowing to add client interfaces sumption. For that purpose, the transponder must possess
only at the network end-points, regardless of service type [3]. cognitive properties in the sense that it should be capable
 2

compensation in single and multiple carrier CPM-capable

Traffic rate [Gbps]

70
60 superchannel-based transponders; and in Section V we eval-
50 uate the power efficiency of the CPM-supported transmission
40 schemes in single and multiple carrier cases, introducing
30 Pareto-efficiency analysis. We also address optimization of
20 the traffic load distribution between the transponder subcar-
10 riers. We propose to employ the minimal rate transmission
0 as a sleep-mode for non-active subcarriers, so that – while
08 12 16 20 00 04 08 substantially reducing the consumption – there is no discon-
Hour tinuity in clock, state of polarization and phase tracking, and
(a) Daily traffic pattern. the subcarrier can be smoothly reactivated. In Section VI
we apply the complexity and optimization analyses to show
by example that in a typical multiple carrier transmission
Traffic rate [Gbps]

70
60 Weekend scenario CPM provides a significant reduction in power
consumption. Finally, in Section VII, we conclude the paper.
50
40 II. B ASIC A SSUMPTIONS
30 For our analysis we make the following assumptions:
20
• Fixed forward error correction (FEC) overhead size:
10
several recent works, e.g., [9], [10], have explored cod-
0
20 21 22 23 24 ing optimization and variable coding rate. Analysis of
Date variable coding rates exceeds the scope of our work. In
(b) Weekly traffic pattern.
particular, we consider a 10−3 pre-FEC BER, required
for the 10−15 pos-FEC BER using a 7% overhead.
Fig. 1: Fluctuation of the aggregate internet traffic between • Maximum symbol rate, Rs
max
, of 28 Gbaud: this rate is
autonomous systems networks. widely explored in long-haul transmission because of the
100 Gbps PM-QPSK systems. It allows to produce two
samples per symbol by commercially available analog to
to identify these fluctuations and adjust its transmission digital converters. We assume that much of the 100G in-
scheme for power efficiency. frastructure will be also used in superchannel systems.
Despite the benefits of cognition attribution, toggling inter- However, the analysis can be easily extended to other
nal parameters of the electronic devices (such as the internal rates.
max
clock rates) working on the edge of their bandwidth capacity • Discrete symbol rates of the form Rs /2l , l = 0, 1, 2, 3,
appears extremely complex and inserts transition-associated resulting in rates of 3.5, 7, 14 and 28 Gbaud. As shown
distortions. Besides, changes in transmission rates require in Section III, these rates can be seamlessly imple-
repeated system synchronization that, in addition to the mented in the Cognitive Power Management scheme by
implementation complexity, results in data loss. Even so, it is symbol repetition. This item distinguishes the subse-
possible to attribute some adaptivity to optical transponders quent efficiency analysis for CPM-capable transponders
while maintaining their operational parameters unchanged from generic software-defined superchannel transpon-
– to avoid the aforementioned complexity – yet obtaining ders, where transmission rates are not constrained by
a noteworthy power efficiency improvement. To accomplish the above relation. In principle, symbol rates may be fur-
that, we have proposed the Cognitive Power Management ther reduced by choosing l > 3 (to 1.75, 0.875... Gbauds),
(CPM) scheme [8], for transmission rate adjustment ac- however, these rates lead to overly fine granularity, not
cording to network traffic fluctuations. CPM is a technique yielding much advantage. Also, these rates may intro-
that allows reducing system power consumption associated duce significant phase noise-associated penalties. This
with chromatic dispersion (CD) compensation. In CPM we is because the variance of the phase noise (modeled as
achieve transmission rate reduction by retransmitting digi- a Wiener process) increases linearly with the sampling
tal symbols, while maintaining internal parameters of the time [11].
electronic devices unchanged. In this way, we avoid the • Reach and the accumulated CD decorrelation: we as-
increased complexity associated with changing the electronic sume a general case where the optical route is not
devices’ operational parameters, as well as the need for known a priori and may contain legacy CD compensa-
repeated synchronization. In [8] we have described CPM for tion modules along some of the sections of the lightpath.
100G systems based on polarization multiplexed (PM) QPSK In this scenario, the accumulated CD ceases to be a
optical transmission. Here, we propose a CPM framework for function of the length of the optical link.
multiple carrier superchannel transmission, extending the We also limit the digital modulation formats to BPSK, QPSK,
previous work. 8QAM and 16QAM. Higher order modulation formats have
The rest of this paper is organized as follows. We begin a prohibitive OSNR penalty that makes their employment
(Section II) by establishing basic assumptions that constitute in this type of systems infeasible.
the framework for the subsequent analysis. Next, in Section
III. C OGNITIVE P OWER M ANAGEMENT
III, we present the operating principle of cognitive power
management scheme in a superchannel systems. We then in- A. System Description
vestigate power consumption related aspects of the transmis- We consider a (multi) terabit long-haul optical link with IP
sion schemes: in Section IV we assess the complexity of CD routers connected to optical line interface by 1, 10, 40 and
 3

CPM- 1G/10G/40G/100G 1G/10G/40G/100G CPM-

Capable Long-Haul Capable
IP GRouter CPM- CPM- IP GRouter
Capable Capable
...

...
Superchannel Superchannel
Line Discrete Line
CPM- 1G/10G/40G/100G Interface transmission rates Interface 1G/10G/40G/100G CPM-
Capable Capable
IP GRouter IP GRouter

Fig. 2: Optical link of a CPM-capable terabit system.

100 Gb/s Ethernet interfaces, as depicted in Fig. 2. The CPM Ts Baudrate

line interface is able to multiplex the incoming information 28G
into a payload of up to the maximum transmission rate and
map into an optical channel for transmission. It is based on Identical symbols
a superchannel optical transponder with N subcarriers. All
14G
subcarriers support QPSK, 8QAM and 16QAM1 polarization-
multiplexed transmission at 28 Gbaud. Also, subcarriers
may be activated separately, so that the number of active Identical symbols
subcarriers takes on values between 1 and N. 7G

B. Operating Principle Identical symbols

CPM is based on the repetition of the transmitted symbols, 3.5G
when a sufficiently low traffic flow is detected. The number of
repetitions per symbol, NR = 2k , k = 0, 1, 2, 3, results in per
subcarrier nominal transmission rates of 200/NR , 150/NR Fig. 3: Repetition-based per subcarrier transmission scheme.
and 100/NR for 16QAM, 8QAM and QPSK respectively. We
consider only homogeneous transmission, where the same
scheme (modulation format and transmission rate) is used synchronized. Thus, a cognitive power management scheme
by all active subcarriers. Combinations of the number of ac- may be resumed in the following steps:
tive subcarriers, modulation formats and the per subcarrier 1) IP routers switch operational rates of their Ethernet
rates result in several transmission schemes, out of which interfaces according to the internal buffer size and his-
only the ones that yield either reach or power consumption torical traffic patterns, through auto-negotiation with
advantages (i.e., Pareto-efficient schemes; see Section V) are the line interface.
supported. For example, considering maximum rate of 1 Tbps 2) CPM line interface monitors the rate of its Ethernet
and a 5-carrier superchannel, transmission rates supported interfaces, computes the optimal rate and adjusts the
by the CPM interface are: 12.5, 18.7, 25, 37.5, 50, 56.2, 62.5, transmission scheme.
75, 93.7, 100, 112.5, 125, 150, 187.5, 200, 225, 250, 300, 375, 3) When the sum of the traffic from all connected IP
400, 450, 500, 600, 750, 800 Gbps and 1 Tbps, based on traffic routers is such that the transmission scheme switch
demands. This granularity may be reduced for simplicity. At at the optical transmitter is triggered, the transmitter
the receiver, the sequence is downsampled according to the receiver synchronization is performed.
required data rate, as shown in Fig. 3. The figure depicts a 4) At the receiver, the downsampling pattern, the number
transmission of two samples per symbol (black arrows), with of active subcarriers, the modulation format and the
Ts being the symbol period. After sampling and analog-to- electronic CD compensation filter length are readjusted
digital (AD) conversion, downsampling is performed, discard- according to the new transmission rate.
ing the samples crossed in red, maintaining two samples for
each transmitted symbol. In this way, system parameters, C. Additional Considerations
such as the transmitter modulator and the receiver AD There are some additional considerations regarding CPM
converters operation rates, remain unchanged. implementation that are left for further study:
In order to benefit from the CPM scheme, IP routers should
• Return-to-zero (RZ) clock synchronization: for RZ sys-
be able to switch the transmission rates of their Ethernet
tems, CPM may require special care during the clock
interfaces between 1, 10, 40 and 100 Gb/s, according to traffic
recovery, or, perhaps, a pulse-carver frequency adjust-
conditions. A decision may be taken based on a composition
ment, so that the signal sampling at the receiver is not
of two parameters: internal buffer congestion and histori-
performed at near zero amplitude instants.
cal traffic patterns. The CPM line interface, in turn, must
• Filter band adjustment: to maximize the signal to noise
monitor the rate of its (client) Ethernet interfaces, compute
ratio (SNR), optical receivers employ electrical filters,
the optimal transmission rate and adjust the multiplexing
whose passband is about 0.5Bs (for coherent receivers)
scheme. It also must acknowledge the receiver at each switch
[12], where Bs is the effective baseband signal band-
of the transmission scheme, so that both line interfaces are
width. Since CPM changes the signal bandwidth, some
1 BPSK modulation is not used because it is not Pareto-efficient in filter adjustment is needed in order to maintain SNR at
CPM (see Section V). its maximum.
 4

• Convergence due to modulation switching. Step 4 of requires three real multiplications and three real additions,
CPM (Section III-B) takes into account online modu- implemented in the following manner [18]: if z1 = a + jb
lation format adjustment. However, its employment is and z2 = c + jd are two complex numbers, then:
not trivial, because it may require repetitive equalizer
z1 × z2 = (ac − bd) + j(ad + bc) (2)
convergence for each modulation switch, which would
add to the latency associated with the CPM synchroniza- = [(a − b)d + a(c − d)] + j [(a − b)d + b(c + d)] . (3)
tion protocols, and perhaps result in data loss. Systems
Because of the identical terms, the computation in (3) re-
that set their modulation format only once, during the
quires three real multiplication and five real additions. Yet,
initialization, according to the estimated OSNR, appear
if the multiplicand, z2 , is known a priori, the sum and
less complex.
the difference of c and d may be pre-computed and stored,
• Anti-bouncing mechanism: since the traffic from the con-
sparing two real additions. This assumption is valid for FFT,
nected routers triggers the transmission scheme switch,
where the multiplicands are of the form ǫ2πkn/N , and is also
threshold values are likely to be employed. This may
valid for the multiplication by the FFT-ed filter coefficients,
lead to back and forth switching (bouncing) caused by
because of the deterministic nature of the chromatic disper-
the ripple in traffic values near the established thresh-
sion. We further assume radix-2 Cooley-Tukey algorithm,
olds. Thus, some mechanism is required to avoid this
because it can be implemented for any power of two FFT
phenomenon.
size, yielding a desired flexibility for rate-adaptive schemes.
Under these assumptions, the number of (non-trivial) real
IV. C OMPUTATIONAL C OMPLEXITY OF THE C HROMATIC multiplications, MR , for FFT of size NF can be computed as
D ISPERSION C OMPENSATION F ILTERS [18]:
A. Per Bit Computational Complexity 3
MR (NF ) = NF (−3 + log2 NF ) + 6, (4)
As was shown in [7], the length (in taps) of the chromatic 2
dispersion compensation bulk filter, NCD , can be computed provided that NF = 2k , k ∈ N. A number of real multiplica-
by the following expression: tions per filtering cycle is given, therefore, by:
 
CDacc λ2 MCY C = p [2MR + 3NF ] (5)
NCD = 2 × + 1, (1)    
2 c Tsp
2
3
= p 2 NF (−3 + log2 NF ) + 6 + 3NF , (6)
where CDacc is the accumulated CD, λ is the optical carrier 2
wavelength, Tsp is the sampling time and c is the speed where p is the number of polarizations. In (5-6) MR is mul-
of light. It was also shown that a 40% reduction in filter tiplied by 2 to account both for the FFT and the subsequent
size, computed by (1), leads to insignificant performance IFFT, and NF is multiplied by 3, as three real multiplications
penalties. Fig. 4 indicates the number of filter taps required are required for each complex one. On the other hand, the
for the CD compensation, as a function of the accumulated number of bits resulting in one filtering cycle is:
CD and the baud rate. The graphic was produced according
pL
to (1) with 40% reduction, and an oversampling factor of bCY C =
log2 M, (7)
κ
2 samples per symbol. The dashed red lines correspond
to CPM-supported symbol rates, namely, 3.5, 7, 14 and 28 where L is the number of equalized samples resulting from
Gbaud, and the dot-dashed black lines correspond to some one filtering cycle in one polarization, κ is the oversampling
fixed accumulated CD values. Note that the filter size is a factor and M is the size of the modulation alphabet. Finally,
linear function of the CD, and a quadratic function of the the number of multiplications per transmitted bit, Mb is
symbol rate, not dependent on any particular modulation computed as:
format. This feature plays an important role in the efficiency Mb = MCY C /bCY C
assessment of the distinct modulation formats.    
κ 3
Having obtained the filter size we estimate the compu- = 2 NF (−3 + log2 NF ) + 6 + 3NF .(8)
L log2 M 2
tational complexity associated with the CD compensation.
Our estimation is based on the number of real multiplication Note that Mb does not depend on the number of polariza-
operations, since they consume more power than the addition tions.
operations, regardless of the hardware implementation [13].
We focus on the frequency domain equalization, because of B. FFT Size Optimization for the CD Compensation Filter
its computational efficiency in comparison with the time
In the overlapping methods (either overlap and save or
domain equalization for a sufficient filter length [8] [14]. The
overlap and add), the number of equalized samples resulting
filtering cycle involves performing the fast Fourier transform
from one filtering cycle, L, is related to the FFT size, NF , by
(FFT) over a new-coming block of data, a term-by-term mul-
[16]:
tiplication of the resulting sequence with the one obtained by
the FFT of the filter coefficients and an inverse FFT (IFFT). L = NF − NCD + 1. (9)
Also, long input data sequence requires some overlapping
method. In [15] a 50% overlap was assumed, that is, an FFT In order to obtain the optimal FFT size, for computational
purposes we first consider a parameter N bF that is the
size, NF , of twice the filter length, which is the minimum
bF in
continuous counterpart of NF , and substitute NF by N
FFT size required for the aliasing free equalization [16].
But computational complexity can be significantly reduced (8-9). Thus, we wish to obtain
by optimizing the FFT size according to the filter length, as b opt
N = arg min (Mb ) . (10)
F
shown in [17]. We assume that each complex multiplication bF
N
 5

28 Gbaud

600

500
Filter size [taps]

400

300
14 Gbaud
200

100 7 Gbaud
3.5 Gbaud
0
6
4
2 2 2.5
1 1.5
x 10
4 0 0.5
10
Accumulated CD [ps/nm] x 10
Baudrate
Fig. 4: CD compensation filter size as a function of the accumulated dispersion and the baud rate.

Substituting (9) in (8), taking the derivative of the latest

with respect to N bF and equating it to zero results in a
transcendental equation for N b opt : 21
F
NF = 2NCD= 512
b opt − log N
2N b opt − 4
Real multiplications/bit

b opt +
log2 N F 2 F
= 3 − ln −1
(2). (11) 20
F
Nb opt − NCD + 1
F

Since 19
k−1 b opt k
2 ≤ N F ≤ 2 , k ∈ N, (12)
18
that is, Nb opt is comprised between two subsequent powers
F
of 2, the optimal FFT size, NFopt , may be found as: 17
 k opt
2 , if Mb (2k ) < Mb (2k−1 ); NF = 2048
NFopt = (13)
2k−1 otherwise. 16
Fig. 5 shows the number of multiplications per transmitted
15
bit for a 250 tap CD compensation filter for a 16QAM
3 4
polarization multiplexed signal, as a function of the FFT 10 10
size. Here, FFT size optimization results in roughly 25% FFT size
complexity reduction in comparison with the 50% overlap Fig. 5: FFT size optimization for a 250 taps CD compensation
case. For practical application, the optimum FFT size values filter, for a 16QAM polarization multiplexed signal with
can be pre-computed and stored in hardware: Fig. 6 shows κ = 2. The solid blue line represents the values of the
optimum FFT size values for different filter lengths. continuous variable N b opt , red asterisks represent possible
F
NF values and the green square indicates the optimum FFT
V. E FFICIENCY E VALUATION OF THE T RANSMISSION opt
size, NF .
S CHEMES
A. Single Carrier Case
Different combinations of the modulation formats and for a given target BER that, in turn, defines system’s reach.
symbol rates listed in Section II, result in several bits per The OSNR is related to SNR and SNR per bit, SNRb , by [19]:
second (nominal) transmission rates for a single carrier: 12.5, p Rs Rb
25, 37.5, 50, 75, 100, 150 and 200 Gbps. Each of these OSNR = SNR = SNRb (14)
2Bref 2Bref
rates (with exception of 150 and 200 Gbps) have more than
one possible form of transmission. As an example, Table I In (14), p is the number of polarizations, Rs is the symbol
resumes the resulting schemes for a 50 Gbps transmission. rate, Bref is the reference band (typically 12.5 GHz) and Rb
Here, PM stands for “Polarization multiplexed”, and SP is the transmission bit rate. The OSNR does not depend on
stands for “Single polarization”. Each of the possible trans- the number of polarizations, but only on the transmission
mission schemes has an associated required OSNR (ROSNR) bit rate [19]. For coherent detection, assuming Gray coding,
 6

12
x 10
8192 5
PM−BPSK @ 28 Gbaud
4096

Real multiplications per second

SP−QPSK @ 28 Gbaud
2048
4 PM−QPSK @ 14 Gbaud
1024
SP−16QAM @ 14 Gbaud
512
PM−16QAM @ 7 Gbaud
256 3
FFT size

128
64
32 2
16
8 1
4
2
0
0 100 200 300 400 500 600 700 0 1 2 3 4
CD compensation filter size Accumulated CD [ps/nm] 4
x 10
Fig. 6: Optimal FFT size values. Fig. 7: Computational complexity of the 50 Gbps transmis-
sion schemes.
TABLE I: Possible Formats for a 50 Gbps Transmission
Nominal Rate [Gbps] Transmission Scheme ROSNR [dB] 12
x 10
4
PM-BPSK @ 28 Gbaud 10.3
PM−BPSK @ 28 Gbaud
PM-QPSK @ 14 Gbaud 10.3
Real multiplications per second

3.5 SP−QPSK @ 28 Gbaud

50 SP-QPSK @ 28 Gbaud 10.3
PM−QPSK @ 14 Gbaud
PM-16QAM@ 7 Gbaud 14.0 3
SP−16QAM @ 14 Gbaud
SP-16QAM @ 14 Gbaud 14.0
2.5 PM−16QAM @ 7 Gbaud

2
the following relations are valid [20]:
1 p  1.5 Pareto frontier
Pb (BPSK, QPSK) = erf c SNRb (15)
2 s ! 1
11 3 SNRb
Pb (8QAM) ≈ erf c √ (16)
16 3+ 3 0.5
r !
3 6 SNRb
Pb (16QAM) ≈ erf c , (17) 10 11 12 13 14
8 15 ROSNR
where Pb is the bit error probability and erf c is a compli- Fig. 8: Pareto efficiency for 50 Gbps transmission schemes.
mentary error function. Note that SNRb required for some
target BER depends only on the constellation format. For the
target BER = 10−3 , (15-17) yield [20]: schemes nor BPSK produce any benefit, considering the
SNRb (BPSK, QPSK) = 6.8 dB (18) above-mentioned features. This can be viewed in terms of
Pareto efficiency for transmission schemes. For instance,
SNRb (8QAM) = 9.0 dB (19)
for a fixed accumulated CD value of 34,000 ps/nm that
SNRb (16QAM) = 10.5 dB. (20) corresponds to transmission over 2000 km of a standard
single mode fiber (SSMF) with dispersion coefficient D = 17
The last column of Table I summarizes the ROSNR computed
ps/nm/km we arrange the possible transmission schemes as
according to (14) for the aforementioned BER in 50 Gbps
in Fig. 8. Here, Pareto frontier is formed by PM-QPSK and
transmission.
PM-16QAM. Other schemes do not meet the Pareto efficiency
Next, for a given rate, the transmission schemes are
criterion (i.e., are not beneficial neither in reach nor in
analyzed taking into consideration two features:
power consumption) and should not be employed. Analogous
1) Computational complexity (as a measure of power con- analysis for all the considered transmission rates results in
sumption) for a given accumulated CD value; a set of Pareto efficient schemes, summarized in Table II.
2) ROSNR at a given target BER. Observe that the schemes are also “vertically” efficient, in
Fig. 7 displays the computational complexity of the trans- a sense that higher rate schemes always result in penalty,
mission schemes of Table I. Note that the relative distance either in reach or power consumption, in comparison with
between the curves remains quasi-constant throughout a lower rate schemes.
wide range of the accumulated CD values. This allows to Mapping in Fig. 9 illustrates the rule of transmission
simplify the analysis assuming some fixed accumulated CD,
without loss of generality. From the inspection of Table I 2 Single polarization scheme appears because of the assumed
and Fig. 7 it is evident that neither single polarization minimum symbol rate of 3.5 Gbaud (see Section II).
 7

TABLE II: Pareto efficient schemes for a single carrier transmission.

Nominal Rate [Gbps] Transmission Schemes ROSNR [dB] Complexity (mult/sec) [Tops] Normalized complexity [%]
PM-QPSK @ 3.5 Gbaud 4.3 0.168 4.4
12.5
SP2 -16QAM@ 3.5 Gbaud 8.0 0.084 2.2
PM-QPSK @ 7 Gbaud 7.3 0.558 15.0
25
PM-16QAM @ 3.5 Gbaud 11.0 0.168 4.4
37.5 PM-8QAM @ 7 Gbaud 11.3 0.558 15.0
PM-QPSK @ 14 Gbaud 10.3 1.535 40.0
50
PM-16QAM @ 7 Gbaud 14.0 0.558 15.0
75 PM-8QAM @ 14 Gbaud 14.3 1.535 40.0
PM-QPSK @ 28 Gbaud 13.3 3.839 100.0
100
PM-16QAM @ 14 Gbaud 17.0 1.535 40.0
150 PM-8QAM @ 28 Gbaud 17.3 3.839 100.0
200 PM-16QAM @ 28 Gbaud 20.0 3.839 100.0

Transmission analog and analog-to-digital converters. Moreover, (quasi)-

Rate (nominal) baud rate subcarrier spacing allows efficient spectrum use.
200G (16, 28) Clearly, power consumption of optical superchannels de-
Non applicable schemes pends on implementation. As of today, there are no finalized
standards, and many issues, such as the implementation
Optimal scheme
method, the number of subcarriers, the level of component
150G (8, 28)
integration etc., are subject to research [21]. For example,
Nyquist-WDM superchannels [22] require a distinct laser
Complexity decreases
assembly for each subcarrier. These assemblies typically
100G (4, 28) (16, 14) dissipate several watts in a steady state operation [23],
depending on the temperature. Coherent optical OFDM
75G (8, 14) schemes, on the other hand, use one optical source. However,
Required rate

a recirculating frequency shift loop [24], used for the “optical

50G (4, 14) (16, 7)
comb” generation, employs EDFA (or a pair of EDFAs) that
37.5G (8, 7)
dissipates several dozens of watts [25]; though future optical
25G (16, 3.5) (4, 7)
(16, 3.5)
integration may substantially lower these values. Same de-
12.5G (4, 3.5)
SP sign tradeoffs apply to generation of local oscillators at the
4 6 8 10 12 14 16 18 20 receiver. One subtlety should be noted: at the transmitter,
ROSNR all subcarriers must remain active, even if there is no data
Estimated OSNR [dB] to transmit. This is because power bursts associated with
Fig. 9: Minimum power transmission scheme selection in turning the subcarriers on and off create transients in the
software defined transponders in bandwidth virtualization fiber, degrading performance. Local oscillators, however, may
environment. be turned on and off according to utilization, without af-
fecting neighboring channels. This adds complexity to power
consumption optimization. In duplex design, local oscillators
scheme selection by a CPM-capable single carrier transpon- are generated by splitting the carrier waves. In this way,
der. Blue asterisks that represent different transmission all system’s laser sources remain active and optimization
schemes are being attributed a pair of numbers, (M, Rs ), reduces to CD compensation associated power.
where M is the constellation size and Rs is the symbol rate 2) Pareto Efficiency Assessment: Consider a superchannel
(in Gbaud). Here, due to Pareto efficiency, for the same trans- duplex transceiver with N subcarriers, where each subcar-
mission rate, schemes to the right of the graph (i.e., with rier is capable of transmitting schemes of Table II. We
higher ROSNR) have lower power consumption. Schemes ignore possible OSNR differences – common to generation
whose ROSNR is higher than the estimated OSNR (inside by recirculating loop – between subcarriers within a super-
the vertical shaded area) are discarded. Likewise, rate re- channel. We further ignore OSNR penalties due to subcarrier
quirements given by the traffic state eliminate schemes with spacing. We also assume homogeneous transmission (see
insufficient rate (inside the horizontal shaded area). Among Section III-B). Due to the variable number of subcarriers,
the remaining schemes (inside the unshaded area), the one new transmission schemes are formed. Table III resumes
of lowest rate and higher ROSNR (right bottom corner of the possible schemes for 400 Gbps transmission for N = 10. Note
unshaded area) is the most power efficient. that due to CPM granularity, not all available subcarriers
can be used. Neglecting OSNR penalties due to subcarrier
spacing, the ROSNR for multicarrier scheme (third column)
B. Multiple Carrier Case is the same as for single carrier with identical modulation
1) General Considerations: Optical superchannels [5], format. In this case the two Pareto efficient schemes are 8
will, most likely, become a new standard for terabit trans- subcarriers PM-QPSK at 14 Gbaud and 8 subcarriers PM-
mission. This is because optical parallelism circumvents
band limitations of the electronic devices, namely, digital-to- 3 Transmission at 450 Gbps.
 8

TABLE III: Possible Formats for 400 Gbps Transmission by this influence reduces, essentializing optical parallelism to
a 10 Subcarrier Superchannel reduce the baud rate and comply with the reach require-
Active Transmission ROSNR Mult./sec. ments.
Subcarriers Scheme [dB] [Tops]
2 PM-16QAM@ 28 Gbaud 20.0 7.678
40 40
3 PM-8QAM3 @ 28 Gbaud 17.3 11.517

Number of subcarriers

Number of subcarriers
1Tbps 1Tbps
700 Gbps 700 Gbps
4 PM-QPSK @ 28 Gbaud 13.3 15.356 30 30
400 Gbps 400 Gbps
4 PM-16QAM@ 14 Gbaud 17.0 6.140
20 20
6 PM-8QAM3 @ 14 Gbaud 14.3 9.210
8 PM-QPSK @ 14 Gbaud 10.3 12.28 10 10
8 PM-16QAM@ 7 Gbaud 14.0 4.464
0 0
0 0.5 1 0 0.5 1
Fraction of Madd Fraction of Madd

16QAM at 7 Gbaud. Despite the previous result, the number (a) OSNR = 10 dB. (b) OSNR = 13 dB.
of active subcarriers for power-efficient transmission in CPM
40
is not necessarily the maximum available: Fig. 10 illustrates 40

Number of subcarriers
Number of subcarriers
1Tbps 1Tbps
the number of active subcarriers in Pareto-efficient schemes 700 Gbps
30
700 Gbps
30 400 Gbps
for transmission rates up to 2 Tbps, for systems with N = 400 Gbps

10, N = 20 and N = 30. With the exception of low rates, 20 20

10 10

30
0 0
0 0.5 1 0 0.5 1
Fraction of Madd Fraction of Madd
25 (c) OSNR = 16 dB. (d) No OSNR restriction.
Active subcarriers

Fig. 11: Number of system’s subcarriers versus the additional

20
complexity.

15 Hence, we propose to adopt the following subcarrier opti-

mization method for CPM transmitters: during the system
10 initialization OSNR estimation is performed at the receiver
and then some of the system subcarriers are permanently
5
N = 10 deactivated to maximize power efficiency, in a similar way
N = 20 to the described in this section. Remaining subcarriers are
N = 30 activated according to the traffic fluctuations, implementing
0 CPM.
0 0.5 1 1.5 2
Nominal rate [Tbps]
Fig. 10: Number of active subcarriers for different transmis- C. Sleep-Mode Implementation
sion rates. For satisfactory performance, some of the receiver signal
processing blocks must operate continuously, so that there is
the number of active subcarriers for Pareto-efficient schemes no data loss due to convergence and synchronization times.
takes on values between ⌊N/2 + 1⌋ and N. In particular, we refer to the clock recovery, carrier phase
3) Optimization of the Number of Subcarriers: Assume and frequency estimations and tracking of the arbitrary
that each subcarrier – even when inactive – increases the polarization rotation in the fiber. In the signal processing
power budget by some constant amount. To match the power chain, some blocks are interdependent; for example, phase
consumption of CD compensation, we analyze this power in- recovery requires a previous CD compensation and polar-
crease in terms of additional complexity. Our reference, Madd ization demultiplexing. Therefore, if completely deactivated,
is the number of real multiplication required for the compen- the subcarriers must be resynchronized upon reactivation.
sation of 34,000 ps/nm of accumulated CD in a single carrier To tackle this problem in CPM transponders, we propose to
PM-QPSK system transmitting at 112 Gbps: 3.839 Tops. (see use the minimum power transmission schemes as the sleep-
Table II). Fig. 11 shows the optimal number of subcarriers for mode for inactive subcarrier instead of a full deactivation.
selected transmission rates, namely, 400 Gbps, 700 Gbps and For example, we can use QPSK at 3.5 Gbaud, whose com-
1 Tbps, where we varied the number of subcarriers between 5 putational complexity for accumulated CD of 34,000 ps/nm
and 45. Each subcarrier added a complexity that is a fraction is 0.17 Tops per second. While not optimal, still, it offers
of Madd , varied in steps of 0.1. We also explored the influence about 96% complexity reduction due to CD compensation in
of the estimated OSNR by repeating the analysis for OSNR comparison with the 28 Gbaud QPSK, whose complexity is
= 10, 13 and 16 dB (Figs. 11a-c), comparing the results with 3.84 Tops per second. In this way, there is no discontinuity
the case of no OSNR restriction (Fig. 11d). If OSNR is high, in clock, state of polarization and phase tracking, and the
even for small additional complexity per subcarrier (10-20% subcarrier can be seamlessly reactivated. The sleep-mode
of CD compensation complexity), transponders with fewer associated additional complexity can be treated as Madd in
subcarriers become more power efficient. As OSNR degrades Section V-B3.
 9

VI. C ASE S TUDY

To exemplify the power efficiency improvement by CPM 1000
we consider a 2000 km uncompensated link of SSMF fiber 6SC PM−[email protected]
6SC PM−QPSK@7Gbaud
with dispersion parameter D = 17 ps/nm/km. The system
transmits at a maximum nominal rate of 1 Tbps over central 800
10SC PM−QPSK@28Gbaud
wavelength λ = 1550 nm using a 10-subcarrier super-

Traffic [Gbps]
channel. The system supports QPSK, 8QAM and 16QAM 600
modulation formats. We further consider that estimated 9SC PM−8QAM@14Gbaud
OSNR = 17 dB and that additional power requirement per 7SC PM−QPSK@28Gbaud
subcarrier is 0.2Madd (see Section V-B). Also, assume an 400
10SC PM−16QAM@7Gbaud
additional OSNR margin of 2 dB. To assess the performance 10SC PM−QPSK@14Gbaud
we scaled the daily traffic pattern of Fig. 1a so that the
200 Traffic profile
maximum rate is 1 Tbps. Fig. 12a exhibits the 24-hour
Transmission rate all formats
CPM transmission pattern of a system represented by the
Transmission rate QPSK
red solid line, where the transponders switch between the 0
modulation formats during operation. Here, the complexity 0 2 4 6 8
is reduced by 36%, from 3.98 × 106 to 2.54 × 106 Tops. The top Time [sec.] 4
x 10
lines of the text boxes denote selected transmission schemes (a) Multiple subcarriers.
for this case. A simpler case – for QPSK modulation only
– is represented by the green dash-dotted line. Note that,
although the transmission rate granularity remains almost
identical, the transmission schemes vary (compare top and 100
bottom lines of the text boxes). For the QPSK modulation
case the complexity is reduced by 28%, from 3.98 × 106 to
2.87 × 106 Tops. 80 PM−QPSK@28Gbaud
Traffic [Gbps]

To illustrate the granularity improvement in multiple-

carrier scenario, we also include a 24-hour CPM transmis-
60
sion pattern for the same link conditions for a single-carrier
PM-QPSK transmission, where the daily traffic pattern is
scaled to the maximum rate of 100 Gbps (Fig. 12b). Here, the PM−QPSK@14Gbaud
40
complexity was reduced by 20%, from 3.98 × 105 to 3.19 × 105
PM−QPSK@7Gbaud
Tops. Complexity reduction in a single carrier CPM, however,
heavily depends on the peak traffic [8]. For example, if the Traffic profile
20
daily traffic pattern of Fig. 1a is not scaled, for the same Transmission rate
conditions, the complexity is reduced by 33% ( from 3.98×105
0 2 4 6 8
to 2.67 × 105 Tops).
Time [sec.] 4
Estimating the associated reduction in power consumption x 10
is a complex task affected by many factors, e.g., data rate, (b) Single carrier.
technology, architecture, packaging, interconnect design, etc. Fig. 12: CPM transmission profile for daily traffic fluctua-
For a gross estimation, however, we used the result of [8] tions.
where we assumed the energy consumption per real mul-
tiplication for 40 nm CMOS ASIC (that includes associated
interconnect capacitance, addition and storage operations) to
be ≈ 2 pJ. Daily power consumption values for the multiple- of Pareto-efficiency regarding transmission rate, reach and
carrier CPM systems are summarized in Table IV. power consumption. We showed that for the same transmis-
sion rate polarization multiplexed and higher modulation
TABLE IV: Daily Power Consumption for Non-CPM-capable, order schemes are more power efficient, and that BPSK
All Modulation Formats and QPSK Only Multiple-Carrier modulation does not yield any advantage – neither in reach
Systems. nor in power efficiency – in CPM-capable systems. We spec-
ified the rule of transmission scheme selection by a CPM-
System Multiplications [Tops] Consumption [kW.Hr]
capable transponder, according to channel conditions. We
Non-CPM 3.98 × 106 2.21 also investigated the optimal number of active subcarriers,
All Formats 2.54 × 106 1.41 showing that Pareto-efficient schemes, for specific transmis-
QPSK 2.87 × 106 1.59 sion maintain active over half the subcarriers. We further
showed the influence of additional complexity on the optimal
degree of system’s optical parallelism, showing that this
influence increases with OSNR. As a consequence of these
VII. C ONCLUSION findings, we proposed to optimize the number of subcarriers
We presented a scheme for the cognitive management for each link during the system initialization, after OSNR es-
of power consumption associated with CD compensation, timation is performed. The transponder would permanently
in software-defined, superchannel optical transponders. The deactivate some subcarriers, so that power consumption is
CPM is based on the intrinsic time variability of inter- minimized. We also proposed to implement subcarrier sleep-
net traffic. We investigated transmission schemes in terms mode by minimum rate transmission, namely, QPSK at 3.5
 10

Gbaud. In this way, there is no discontinuity in clock, state [16] J. Proakis and D. Manolakis, Digital signal processing: prin-
of polarization and phase tracking, and the subcarrier can ciples, algorithms, and applications, 3rd ed. Prentice Hall,
1996.
be smoothly reactivated when required. Finally, we demon-
[17] J. C. Geyer, C. R. Fludger, T. Duthel, C. Schulien, and
strated the reduction in power consumption through a case B. Schmauss, “Efficient frequency domain chromatic dispersion
study of a typical terabit transmission scenario, showing a compensation in a coherent polmux qpsk-receiver,” in Optical
36% reduction in an ideal case (using all modulation formats) Fiber Communication Conference. Optical Society of America,
and 28% reduction in the QPSK only case. The 8% difference 2010, p. OWV5.
[18] R. Blahut, Fast Algorithms for Signal Processing. Cambridge
between both cases that corresponds to non-negligible 180 University Press, 2010.
W.Hr of daily consumption suggests to investigate a seam- [19] R.-J. Essiambre, G. Kramer, P. Winzer, G. Foschini, and
less online toggle between modulation formats in software- B. Goebel, “Capacity limits of optical fiber networks,” Lightwave
defined transponders. Technology, Journal of, vol. 28, no. 4, pp. 662 –701, feb.15, 2010.
[20] E. Ip, A. P. T. Lau, D. J. F. Barros, and J. M. Kahn, “Coherent
detection in optical fiber systems,” Opt. Express, vol. 16, no. 2,
A CKNOWLEDGMENT pp. 753–791, Jan 2008.
[21] “Super-channels DWDM transmission beyond 100Gb/s,” Infin-
This work was supported by the Innovation Center, Eric- era Corporation, 2011, White paper, Doc. number WP-SC-12-
sson Telecomunicaes S.A., Brazil. 2011.
[22] G. Bosco, A. Carena, V. Curri, P. Poggiolini, and F. Forghieri,
“Performance limits of Nyquist-WDM and CO-OFDM in high-
R EFERENCES speed PM-QPSK systems,” Photonics Technology Letters, IEEE,
vol. 22, no. 15, pp. 1129 –1131, aug.1, 2010.
[1] C. Glingener, “Optical networking trends and evolution,” in [23] “ITLA-3020-C / ITLA-3010-L, MSA-compliant integrated tun-
Optical Fiber Communication Conference. Optical Society of able laser assembly,” Santur Corporation, 2006, Product Data
America, 2011, p. OThAA1. Sheet, Doc. ITLA-30X0-Y.
[2] “Spectral grids for WDM applications: DWDM frequency grid. [24] Y. Ma, Q. Yang, Y. Tang, S. Chen, and W. Shieh, “1-Tb/s
Recommendation ITU-T G.694.1,” Telecommunication Stan- per channel coherent optical OFDM transmission with sub-
dardization Sector of International Telecommunication Union, wavelength bandwidth access,” in Conference on Optical Fiber
Feb. 2012. Communication , 2009. OFC 2009. Optical Society of America,
[3] “Bandwidth virtualization enables a programmable optical net- 2009, pp. 1 – 3.
work,” Infinera Corporation, 2007, White paper, Doc. number [25] “Gain flattened inline EDFA for DWDM networks,” Optilab,
094-002-0108. 2010, Product Data Sheet.
[4] M. Eiselt, B. Teipen, K. Grobe, A. Autenrieth, and J.-P. Elbers,
“Programmable modulation for high-capacity networks,” in Op-
tical Communication (ECOC), 2011 37th European Conference
and Exhibition on, sept. 2011, pp. 1 –3.
[5] G. Bosco, V. Curri, A. Carena, P. Poggiolini, and F. Forghieri,
“On the performance of Nyquist-WDM terabit superchannels
based on PM-BPSK, PM-QPSK, PM-8QAM or PM-16QAM sub-
carriers,” Lightwave Technology, Journal of, vol. 29, no. 1, pp.
53 –61, jan.1, 2011.
[6] “Center of study and research in network technology and
operations CEPTRO,” available at: http://www.ptt.br/.
[7] S. J. Savory, “Digital filters for coherent optical receivers,”
Optics Express, vol. 16, no. 2, pp. 804–817, 2008.
[8] V. Rozental, G. Bruno, G. Lombardi, R. Grosso, A. Alping,
and D. Mello, “Cognitive power management in 100G optical
transponders,” in International Conference on Transparent Op-
tical Networks (ICTON), 2012, July 2012.
[9] S. Desbruslais and S. Savory, “Optimizing the fec overhead
in a 100 gbe pdm-qpsk digital coherent transmission system,”
Lightwave Technology, Journal of, vol. 29, no. 8, pp. 1118 –1126,
april15, 2011.
[10] G.-H. Gho and J. Kahn, “Rate-adaptive modulation and coding
for optical fiber transmission systems,” Lightwave Technology,
Journal of, vol. 30, no. 12, pp. 1818 –1828, june15, 2012.
[11] F. A. Garcia, D. A. Mello, and H. Waldman, “Feedforward
carrier recovery for polarization demultiplexed signals with
unequal signal to noise ratios,” Opt. Express, vol. 17, no. 10,
pp. 7958–7969, May 2009.
[12] A. Carena, V. Curri, G. Bosco, P. Poggiolini, and F. Forghieri,
“Modeling of the impact of nonlinear propagation effects in un-
compensated optical coherent transmission links,” Lightwave
Technology, Journal of, vol. 30, no. 10, pp. 1524 –1539, may
2012.
[13] A. Faria, L. de Aguiar, D. Lara, and A. Loureiro, “Compar-
ative analysis of power consumption in the implementation
of arithmetic algorithms,” in Trust, Security and Privacy in
Computing and Communications (TrustCom), 2011 IEEE 10th
International Conference on, nov. 2011, pp. 1247 –1254.
[14] B. Spinnler, “Equalizer design and complexity for digital coher-
ent receivers,” Selected Topics in Quantum Electronics, IEEE
Journal of, vol. 16, no. 5, pp. 1180 –1192, sept.-oct. 2010.
[15] S. Savory, “Digital signal processing for coherent systems,” in
Optical Fiber Communication Conference. Optical Society of
America, 2012, p. OTh3C.7.