Abstract—We extend the previously proposed cognitive have described CPM for 100G systems based on polariza-
power control (CPM) scheme for efficient power manage- tion multiplexed (PM) QPSK optical transmission. Here,
ment in optical transponders, to include multiple carrier
we extend our analysis for multiple carrier transmission
transmission in a bandwidth virtualization environment.
The scheme reduces the length of electronic chromatic in a bandwidth virtualization environment. In particular,
dispersion compensation filters according to the network we perform a complexity analysis of CD compensation for
traffic fluctuations. Transmission rate reduction is achieved possible transmission schemes in single and multiple car-
by symbol retransmission, while maintaining internal pa- rier software-defined transponders, introducing the notion
rameters of the electronic devices unchanged, to avoid
increased complexity. We investigate single and multiple of Pareto-efficiency for optical transmission schemes. We
carrier transmission schemes in terms of Pareto-efficiency further address the issue of optimizing the traffic load dis-
regarding transmission rate, reach and power consumption. tribution between the subcarriers in a CPM superchannel.
We show that for the same transmission rate polarization Finally, we show by example that in a typical multiple carrier
multiplexed and higher modulation order schemes are more
power efficient. We further investigate the optimal number
transmission scenario CPM provides a significant reduction
of active subcarriers in a CPM-capable superchannel-based in power consumption.
optical line interface, showing that schemes that keep active The rest of this paper is organized as follows: Section II
over half the subcarriers are power efficient. We realize a analyzes the computational complexity of CD compensation;
case study to demonstrate power savings offered by CPM in Section III assesses the power efficiency of the transmission
a typical terabit transmission scenario.
schemes in single carrier and multiple carrier cases; Section
Index Terms—Optical transponders; bandwidth virtualiza- IV describes the cognitive power management scheme in a
tion; cognitive power management; chromatic dispersion
superchannel environment; Section V presents a case study
compensation; Pareto-efficient transmission.
of power consumption for a typical transmission scenario and
Section VI concludes the paper.
I. I NTRODUCTION
II. C OMPUTATIONAL C OMPLEXITY OF THE C HROMATIC
ommercial product development for 100G optical sys-
C tems raised an important issue of power consumption
in optical transponders, leading to such solutions as single
D ISPERSION C OMPENSATION F ILTERS
As was shown in [4], the length (in taps) of the chromatic
dispersion compensation bulk filter, NCD , can be computed
chip ADC/DSP, use of advanced CMOS processes and optical
by the following expression:
integration. Higher per channel transmission rates like 400
Gbps and 1 Tbps will rely, at least in the near future, on CDacc λ2
NCD = 2 × + 1, (1)
optical parallelism that allows to overcome electronic de- 2 c Tsp
2
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. 1: CD compensation filter size as a function of the accumulated dispersion and the baud rate.
We focus on the frequency domain equalization, because of provided that NF = 2k , k ∈ N. A number of real multiplica-
its computational efficiency in comparison with the time tions per filtering cycle is given, therefore, by:
domain equalization for a sufficient filter length [3] [6]. The
filtering cycle involves performing the fast Fourier transform MCY C = p [2MR + 3NF ] (5)
(FFT) over a new-coming block of data, a term-by-term mul- 3
= p 2 NF (−3 + log2 NF ) + 6 + 3NF , (6)
tiplication of the resulting sequence with the one obtained by 2
the FFT of the filter coefficients and an inverse FFT (IFFT).
where p is the number of polarizations. In (5-6) MR is mul-
Also, long input data sequence requires some overlapping
tiplied by 2 to account both for the FFT and the subsequent
method. In [7] a 50% overlap was assumed, that is, an FFT
IFFT, and NF is multiplied by 3, as three real multiplications
size, NF , of twice the filter length, which is the minimum
are required for each complex one. On the other hand, the
FFT size required for the aliasing free equalization [8]. But
number of bits resulting in one filtering cycle is:
computational complexity can be significantly reduced by
optimizing the FFT size according to the filter length, as pL
bCY C = log2 M, (7)
shown in [9]. For our analysis we assume that each complex κ
multiplication requires three real multiplications and three where L is the number of equalized samples resulting from
real additions, implemented in the following manner [10]: if one filtering cycle in one polarization, κ is the oversampling
z1 = a + jb and z2 = c + jd are two complex numbers, then: factor and M is the size of the modulation alphabet. Finally,
the number of multiplications per transmitted bit, Mb is
z1 × z2 = (ac − bd) + j(ad + bc) (2) computed as:
= [(a − b)d + a(c − d)] + j [(a − b)d + b(c + d)] . (3)
Mb = MCY C /bCY C
Because of the identical terms, the computation in (3) re- κ 3
= 2 NF (−3 + log2 NF ) + 6 + 3NF .(8)
quires three real multiplication and five real additions. Yet, L log2 M 2
if the multiplicand, z2 , is known a priori, the sum and
Note that Mb does not depend on the number of polariza-
the difference of c and d may be pre-computed and stored,
tions. In the overlapping methods (either overlap and save or
sparing two real additions. This assumption is valid for FFT,
overlap and add), the number of equalized samples resulting
where the multiplicands are of the form ǫ2πkn/N , and is also
from one filtering cycle, L, is related to the FFT size, NF , by
valid for the multiplication by the FFT-ed filter coefficients,
[8]:
because of the deterministic nature of the chromatic disper-
sion. We further assume radix-2 Cooley-Tukey algorithm, L = NF − NCD + 1. (9)
because it can be implemented for any power of two FFT
size, yielding a desired flexibility for rate-adaptive schemes. In order to obtain the optimal FFT size, for computational
Under these assumptions, the number of (non-trivial) real purposes we first consider a parameter N bF that is the
multiplications, MR , for FFT of size NF can be computed as bF in
continuous counterpart of NF , and substitute NF by N
[10]: (8-9). Thus, we wish to obtain
3 b opt
N = arg min (Mb ) . (10)
MR (NF ) = NF (−3 + log2 NF ) + 6, (4) F
bF
N
2
3
8192
21 4096
NF = 2NCD= 512 2048
Real multiplications/bit
20 1024
512
256
FFT size
19
128
18 64
32
17 16
8
Nopt
F
= 2048
16 4
2
15 0 100 200 300 400 500 600 700
3 4 CD compensation filter size
10 10
FFT size Fig. 3: Optimal FFT size values.
Fig. 2: FFT size optimization for a 250 taps CD compensation
filter, for a 16QAM polarization multiplexed signal with κ =
2. The solid blue line represents the values of the continuous 100 Gbps PM-QPSK systems. It allows to produce two
variable N b opt , red asterisks represent possible values for NF samples per symbol by commercially available analog to
F
and the green square indicates the optimum FFT size, NFopt . digital converters.
max
• Discrete symbol rates of the form Ts /2l , l = 0, 1, 2, 3,
resulting in rates of 3.5, 7, 14 and 28 Gbaud. As shown in
Substituting (9) in (8), taking the derivative of the latest Section IV, these rates can be seamlessly implemented
with respect to N bF and equaling it to zero results in a in the Cognitive Power Management scheme by symbol
transcendental equation for N b opt : repetition.
F
• Reach and the accumulated CD decorrelation: since
b opt b opt
b opt + 2NF − log2 NF − 4
log2 N = 3 − ln−1 (2). (11)
we consider a bandwidth virtualization scenario, the
F
Nb opt − NCD + 1 service layer is decoupled from the transmission layer
F
[1]. Therefore, the optical route is not known a priori
Since
and may or may not contain legacy CD compensation
2k−1 ≤ N
b opt ≤ 2k ,
F k ∈ N, (12) modules along some or all the sections of the lightpath.
In this scenario, the accumulated CD ceases to be a
that is, Nb opt is comprised between two subsequent powers
F function of the length of the optical link.
of 2, the optimal FFT size, NFopt , may be found as:
k We also consider a limited set of digital modulation formats,
2 , if Mb (2k ) < Mb (2k−1 ); namely, BPSK, QPSK, 8QAM and 16QAM. Higher order
NFopt = k−1 (13)
2 otherwise. modulation formats have a prohibitive OSNR penalty that
makes their employment in this type of systems infeasible.
Fig. 2 shows the number of multiplications per transmitted
The aforementioned modulation formats, in combination
bit for a 250 tap CD compensation filter for a 16QAM
with the considered symbol rates, result in several bits per
polarization multiplexed signal, as a function of the FFT
second (nominal) transmission rates for a single carrier: 12.5,
size. Here, FFT size optimization results in roughly 25%
25, 37.5, 50, 75, 100, 150 and 200 Gbps. Each of these rates
complexity reduction in comparison with the 50% overlap
(with exception of 150 and 200 Gbps) have, essentially, more
case. For practical application, the optimum FFT size values
than one possible form of transmission. As an example, Table
can be pre-computed and stored in hardware, as can be seen
I resumes the resulting schemes for a 50 Gbps transmission.
in Fig. 3, that shows optimum FFT size values for different
Here, PM stands for “Polarization multiplexed”, and SP
filter lengths.
TABLE I: Possible Formats for a 50 Gbps Transmission
III. E FFICIENCY E VALUATION OF THE T RANSMISSION
S CHEMES Nominal Rate [Gbps] Transmission Scheme ROSNR [dB]
12 12
x 10 x 10
5 4
PM−BPSK @ 28 Gbaud PM−BPSK @ 28 Gbaud
Real multiplications per second
2
2
1.5 Pareto frontier
1
1
0.5
0
0 1 2 3 4 10 11 12 13 14
Accumulated CD [ps/nm] 4
ROSNR
x 10
Fig. 4: Computational complexity of the 50 Gbps transmis- Fig. 5: Pareto efficiency for 50 Gbps transmission schemes.
sion schemes.
power.
50G (4, 14) (16, 7)
Consider a superchannel duplex transceiver with N sub-
37.5G (8, 7)
carriers, where each subcarrier is capable of transmitting
25G (16, 3.5) (4, 7)
(16, 3.5) schemes of Table II. We ignore possible OSNR differences
12.5G (4, 3.5)
SP – common to generation by recirculating loop – between
4 6 8 10 12 14 16 18 20 subcarriers within a superchannel. We further ignore OSNR
ROSNR penalties due to subcarrier spacing, assuming that all
Estimated OSNR [dB] schemes are equally affected. We also assume homogeneous
Fig. 6: Minimum power transmission scheme selection in transmission, that is, same scheme is used by all (active)
software defined transponders in bandwidth virtualization subcarriers. Due to the variable number of subcarriers, new
environment. transmission schemes are formed. Table III resumes possible
schemes for 400 Gbps transmission for N = 10. Note that due
Number of subcarriers
Number of subcarriers
1Tbps 1Tbps
available: Fig. 7 illustrates the number of active subcarriers 700 Gbps 700 Gbps
30 30 400 Gbps
in Pareto-efficient schemes for transmission rates up to 2 400 Gbps
10 10
30
0 0
0 0.5 1 0 0.5 1
Fraction of Madd Fraction of Madd
25
(a) OSNR = 10 dB. (b) OSNR = 13 dB.
Active subcarriers
20 40 40
Number of subcarriers
Number of subcarriers
1Tbps 1Tbps
700 Gbps 700 Gbps
30 30
400 Gbps 400 Gbps
15
20 20
10 10 10
N = 10 0 0
5 0 0.5 1 0 0.5 1
N = 20 Fraction of Madd Fraction of M
add
N = 30 (c) OSNR = 16 dB. (d) No OSNR restriction.
0
0 0.5 1 1.5 2 Fig. 8: Number of system’s subcarriers versus the additional
Nominal rate [Tbps]
complexity.
Fig. 7: Number of active subcarriers for different transmis-
sion rates.
Traffic rate [Gbps]
70
transmission rates, namely, 400 Gbps, 700 Gbps and 1 Tbps, 60 Weekend
where we varied the number of subcarriers between 5 and 50
45. Each subcarrier added a complexity that is a fraction of 40
Madd , varied in steps of 0.1. We also explored the influence 30
of the estimated OSNR by repeating the analysis for OSNR 20
= 10, 13 and 16 dB (Figs. 8a-c), comparing the results with 10
the case of no OSNR restriction (Fig. 8d). If OSNR is high, 0
20 21 22 23 24
even for small additional complexity per subcarrier (10-20% Date
of CD compensation complexity), transponders with fewer
subcarriers become more power efficient. As OSNR degrades (b) Weekly traffic pattern.
this influence reduces, essentializing optical parallelism to Fig. 9: Fluctuation of the aggregate internet traffic between
reduce the baud rate and comply with the reach require- ASNs.
ments.
...
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
B. Operating Principle red, maintaining two samples for each transmitted symbol.
We consider a terabit long-haul optical link with IP routers In this way, system parameters, such as the transmitter
connected to optical line interface by 1, 10, 40 and 100 Gb/s modulator and the receiver AD converters operation rates,
Ethernet interfaces, as depicted in Fig. 10. The superchannel remain unchanged. For RZ systems this scheme may require
based CPM line interface is able to multiplex the incoming special care during the clock recovery, or, perhaps, a pulse-
information into a payload of up to the maximum trans- carver frequency adjustment and is left for further study.
mission rate and map into an optical channel for trans- In order to benefit from the CPM scheme, IP routers should
mission. All subcarriers support QPSK, 8QAM and 16QAM be able to switch the transmission rates of their Ethernet
polarization-multiplexed transmission at 28 Gbaud. interfaces between 1, 10, 40 and 100 Gb/s, according to traffic
CPM is based on the repetition of the transmitted symbols, conditions. A decision may be taken based on a composition
when a sufficiently low traffic flow is detected. The number of of two parameters: internal buffer congestion and histori-
repetitions per symbol, NR = 2k , k = 0, 1, 2, 3, results in per cal traffic patterns. The CPM line interface, in turn, must
subcarrier nominal transmission rates of 200/NR , 150/NR monitor the rate of its (client) Ethernet interfaces, compute
and 100/NR for 16QAM, 8QAM and QPSK respectively. As in the optimal transmission rate and adjust the multiplexing
Section III-B, we consider only homogeneous transmission, scheme. It also must acknowledge the receiver at each switch
where the same scheme (modulation format and transmis- of the transmission scheme, so that both line interfaces are
sion rate) is used by all active subcarriers. Combinations synchronized. Thus, a cognitive power management scheme
of the number of active subcarriers, modulation formats may be resumed in the following steps:
and the per subcarrier rates result in several transmission 1) IP routers switch operational rates of their Ethernet
schemes, out of which only the Pareto-efficient ones are interfaces according to the internal buffer size and his-
supported (see Section III-B). For example, considering max- torical traffic patterns, through auto-negotiation with
imum rate of 1 Tbps and a 5-carrier superchannel, transmis- the line interface.
sion rates supported by the CPM interface are: 12.5, 18.7, 25, 2) CPM line interface monitors the rate of its Ethernet
37.5, 50, 56.2, 62.5, 75, 93.7, 100, 112.5, 125, 150, 187.5, 200, interfaces, computes the optimal rate and adjusts the
225, 250, 300, 375, 400, 450, 500, 600, 750, 800 Gbps and 1 transmission scheme.
Tbps, based on traffic demands. Surely, this granularity may 3) When the sum of the traffic from all connected IP
be reduced for simplicity. At the receiver, the sequence is routers is such that the transmission scheme switch
downsampled according to the required data rate, as shown at the optical transmitter is triggered, the transmitter
in Fig. 11. The figure depicts a transmission of two samples receiver synchronization is performed.
4) At the receiver, the downsampling pattern, the number
Ts Baudrate of active subcarriers and the electronic CD compensa-
28G tion filter length are readjusted according to the new
transmission rate.
Identical symbols Ideally, in step 4 the modulation format would also be
14G
adjusted. However, employment of such systems is question-
able because equalizer convergence time in each modula-
tion switch adds to the latency associated with the CPM
Identical symbols synchronization protocols, and may also result in data loss.
7G Therefore, we assume that the system sets its modulation
format only once, during the initialization, according to the
estimated SNR.
Identical symbols
3.5G
V. C ASE S TUDY
To exemplify the power efficiency improvement by CPM
Fig. 11: Repetition-based per subcarrier transmission we consider a 2000 km uncompensated link of SSMF fiber
scheme. with dispersion parameter D = 17 ps/nm/km. The system
transmits at a maximum nominal rate of 1 Tbps over central
per symbol (black arrows), with Ts being the symbol period. wavelength λ = 1550 nm using a 10-subcarrier superchan-
After sampling and analog-to-digital (AD) conversion, down- nel. The system supports QPSK, 8QAM and 16QAM modu-
sampling is performed, discarding the samples crossed in lation formats. We further consider that estimated OSNR =
8
17 dB and that additional power requirement per subcarrier with a single carrier case, we extended our analysis for a
is 0.2Madd (see Section III-B). Also, assume an additional multiple carrier superchannel bandwidth virtualization en-
OSNR margin of 2 dB. To assess the performance we scaled vironment. We investigated transmission schemes in terms
the daily traffic pattern of Fig. 9a so that the maximum rate of Pareto-efficiency regarding transmission rate, reach and
is 1 Tbps. Fig. 12 exhibits the 24-hour CPM transmission power consumption. We showed that for the same transmis-
pattern of an ideal system represented by the red solid sion rate polarization multiplexed and higher modulation
line, where the transponders switch between the modulation order schemes are more power efficient, and that BPSK
formats during operation. Here, the complexity is reduced modulation does not yield any advantage – neither in reach
by 36%, from 3.98 × 106 to 2.54 × 106 Tops. The top lines nor in power efficiency – in CPM-capable systems. We spec-
of the text boxes denote selected transmission schemes for ified the rule of transmission scheme selection by a CPM-
this case. A simpler case – for QPSK modulation only – capable transponder, according to channel conditions. We
is represented by the green dash-dotted line. Note that, also investigated the optimal number of active subcarriers
although the transmission rate granularity remains almost in superchannel-based optical line interface, showing that
identical, the transmission schemes vary (compare top and schemes that keep active over half the subcarriers are power
bottom lines of the text boxes). For the QPSK modulation efficient for some transmission rate. We further showed the
case the complexity is reduced by 28%, from 3.98 × 106 to influence of additional complexity on the optimal degree of
2.87 × 106 Tops. system’s optical parallelism, showing that this influence in-
creases with OSNR. Finally, we demonstrated the reduction
in power consumption through a case study of a typical
terabit transmission scenario, showing a 36% reduction in an
1000
ideal case (using all modulation formats) and 28% reduction
6SC PM−16QAM@3.5Gbaud
6SC PM−QPSK@7Gbaud in the QPSK only case. The 8% difference between both
800 cases that corresponds to non-negligible 180 W.Hr of daily
10SC PM−QPSK@28Gbaud consumption suggests to investigate a seamless online toggle
Traffic [Gbps]