s discrete multi-tone

It’s a preprinted version for researchgate, all rights reserved by Optics Letters, OSA.

Nonlinearity-aware 200-Gbit/s discrete multi-tone transmission for C-band short-reach optical interconnects with a single packaged EML LU ZHANG,1,3 XUEZHI HONG,3,4 XIAODAN PANG,2,3 OSKARS OZOLINS,2 ALEKSEJS UDALCOVS,2 RICHARD SCHATZ,3 CHANGJIAN GUO,4 JUNWEI ZHANG,4 FREDRIK NORDWALL,5 KLAUS M. ENGENHARDT,6 URBAN WESTERGREN,3 SERGEI POPOV,3 GUNNAR JACOBSEN,2 SHILIN XIAO,1 WEISHENG HU,1 JIAJIA CHEN3,4,* 1State

Key Laboratory of Advanced Optical Communication System and Networks, Shanghai Jiao Tong University, Shanghai 200240, China and Transmission Laboratory, RISE Acreo AB, Kista 16425, Sweden 3School of ICT, KTH Royal Institute of Technology, Kista 16440, Sweden 4MOE International Laboratory for Optical Information Technologies, South China Academy of Advanced Optoelectronics, South China Normal University, Guangzhou 510006, China. 5Tektronix AB, Stockholm16432, Sweden 6Tektronix GmbH, Stuttgart 70182, Germany *Corresponding author: [email protected] 2Networking

Received XX Month XXXX; revised XX Month, XXXX; accepted XX Month XXXX; posted XX Month XXXX (Doc. ID XXXXX); published XX Month XXXX

We experimentally demonstrate the transmission of 200Gbit/s discrete multi-tone (DMT) at soft-FEC limit in an intensity-modulation direct-detection system with a single C-band packaged distributed feedback laser and traveling-wave electro absorption modulator (DFBTWEAM), digital-to-analog converter (DAC) and photodiode. The bits and power loaded DMT signal is transmitted over 1.6-km standard single mode fiber (SSMF) with a net rate of 166.7-Gbit/s, achieving an effective electrical spectrum efficiency of 4.93-bit/s/Hz. Meanwhile, net rates of 174.2-Gbit/s and 179.5-Gbit/s are also demonstrated over 0.8-km SSMF and in optical back-to-back case, respectively. The feature of the packaged DFB-TWEAM is presented. The nonlinearityaware digital signal processing algorithm for channel equalization is mathematically described, which improves the signal-to-noise ratio up to 3.5-dB. © 2017 Optical Society of America OCIS codes: (060.2330) Fiber optics communications; (060.4080) Modulation. http://dx.doi.org/10.1364/OL.xxxxxx

With the continuously growing popularity of Internet applications, the overall datacenter traffic has been dramatically increasing since the last decade. It is expected that by 2020 most of the datacenter traffic will stay within the datacenters [1], which

drives the demands for high-capacity short-reach interconnections. The standardization group of next generation short-reach interfaces aims at 200-GbE and 400-GbE [2], in which the data rate per channel is improved from 10- or 25-Gbit/s in the current 100-GbE interface standard [3] to 50-Gbit/s, 100-Gbit/s or even higher. To fulfill the capacity demand while keeping a low implementation cost, optical interconnects based on intensitymodulation direct-detection (IM/DD) system with advanced modulation formats have been intensively investigated. Although there are some works reported in O-band (i.e., 1.3μm), such as net rate 200-Gbit/s 4-level pulse amplitude modulation (PAM-4) transmission [4], net rate 250-Gbit/s discrete multi-tone (DMT) transmission [5], C-band (i.e., 1.55-μm) is considered to be more attractive to provide ultra-high capacity in the magnitude of Tera bits per second and beyond for short-reach communications within datacenters thanks to its maturity in dense wavelength division multiplexing (DWDM) devices. In Cband, several works achieving high data rate have been reported [6-10]. In [6], net rate 140-Gbit/s PAM-4 transmission is carried out up to 1.5-km standard single mode fiber (SSMF), while net rate 186.9-Gbit/s PAM-4 transmission is limited to 0.5-km SSMF. [7] has reported net rate 186.9-Gbit/s PAM-4 and 250-Gbit/s PAM-8 transmission with 44-GHz InP-based distributed feedback laser (DFB)-MZM over 1.2-km SSMF. However, the system capacity is limited by the bandwidth of the modulator. Besides, some works adapt multiplexing techniques to improve system capacity at the expense of more resources, such as dual polarization PAM-4 transmission (net rate 158.7-Gbit/s per polarization) using Stokes-


Rx-DSP

P/S

Demapping

FFT

Equalization

Serial to Parallel

TD-NE

PD

CP Removal

(a)

Synchronization

SSMF EDFA VOA

Downsampling

CP Insertion

Parallel to Serial

IFFT

Bit-Power Loading

Mapping

EA

Serial to Parallel

200GSa/s RTO

100-GHz DFB-TWEAM

92GSa/s DAC

Tx-DSP

(b)

Fig. 1. Experimental setup and DSP flow. (a) packaged DFB-TWEAM, (b) optical spectrum of modulated laser with different modulation amplitudes.

vector receiver [8], and DMT transmission employing four wavelength channels (net rate 128.5-Gbit/s per channel) [9] or two sidebands (127.15-Gbit/s per band) [10]. Nevertheless, the Cband IM/DD system using cost-effective and high-speed packaged optoelectronic devices to reach 200-Gbit/s per wavelength per polarization before forward error correction (FEC) is still challenging. Due to the dispersion induced power-fading at C-band, the useable signal bandwidth is limited even for short-reach communications (e.g., 6-dB bandwidth for 2-km SSMF in double sideband IM/DD systems is ~ 34-GHz). Therefore, improving the system’s spectrum efficiency (SE) is crucial to further enhance Cband IM/DD system capacity. Compared with OOK, PAM-4 and PAM-8, DMT can adaptively allocate different modulation orders and power to orthogonal subcarriers based on their channel status and hence is relatively easy to achieve higher SE. According to the evaluation presented in IEEE802.3bs 400-GbE task force meeting [11], DMT requires smallest quantity of components and the narrowest bandwidth, and thus can be considered as a promising candidate for future 200-GbE and 400-GbE optical interconnects. In this letter, we experimentally demonstrate a 200-Gbit/s DMT modulated IM/DD system using a single packaged 1.55-μm externally modulated laser (DFB laser and traveling-wave electro absorption modulator, DFB-TWEAM, previously used up to 116Gbit/s [12]), a single digital-to-analog converter (DAC), and a single photodiode (PD). The system architecture and digital signal process (DSP) flow are shown in Fig. 1. The DSP flow is similar to that in a conventional DMT system except for a simplified time domain nonlinear equalizer (TD-NE). The bit and power loading with Chow algorithm [13] is utilized at the transmitter to enhance the capacity and SE of the DMT system. The linear equalizer [14] is used to compensate the channel chromatic dispersion and reduce the system additive Gaussian noise influence. The DMT samples are generated offline by MATLAB software and loaded into a 92GSa/s DAC (Keysight AWG M8196A. 3-dB bandwidth: 32-GHz, 8bit vertical resolution). The signal from the DAC is amplified by an electrical amplifier (EA) with 11-dB gain (3-dB bandwidth: 65GHz) before applying on the DFB-TWEAM for modulation. The PIN-PD (3-dB bandwidth: 90-GHz) is a high-speed InP-based O/E converter packaged prototype photodetector from u2t (currently Finisar) with a responsivity of 0.5-A/W. The electrical signal from direct-detection is captured by one 200-GSs/s real-time oscilloscope (RTO, Tektronix DPO77002SX. 3-dB bandwidth: 70GHz). The picture of the packaged DFB-TWEAM is shown in Fig. 1 (a). The device is composed of a monolithically integrated DFB laser with an EAM designed with segmentation method. The DFB gain

section consisted of 7 quantum wells (QWs), and the modulator had 12 QWs of 9-nm thickness. The chip size is 1×0.5-mm2 and the connector is W1 type. The slope efficiency of the modulator is 0.04-W/A, which allows us to reach about 2-mW with only 80-mA driving current. Besides, the device has a static extinction ratio during modulation in the range of 0 to 35-dB, which is related to modulator bias voltage and swing of driving voltage. To balance linearity and modulation depth, the bias point and laser driving current of modulator are set to -1.75V and 122-mA, respectively. Moreover, the device has beyond 100-GHz 3-dB bandwidth with less than 2-dB ripple in the pass band of the EML. However, as Fig. 1 (b) shows, the optical spectrum of DFB-TWEAM with different modulation amplitudes (peak-to-peak voltage) shows a red shift with respect to larger electrical signal amplitude, which is due to the thermal regime variation caused by microwave absorption of impedance matching resistors [15]. This implies that memory nonlinearities are introduced when the DMT signal with a large swing is used for modulation. Besides the modulation nonlinearity of EML, the nonlinearity of the short-reach DMT fiber communication system also comes from other sources, such as inter-subcarrier mixing in square-law detection, clipping of the signal and the nonlinearity from electrical amplifiers due to the high peak-to-average power ratio (PAPR) of the DMT signal. To mitigate such detrimental effects, nonlinearityaware DSP based on TD-NE is utilized. To reduce the DSP complexity, a simplified nonlinear model considering only the 2ndorder and partial 3rd-order terms of the conventional Volterra series model [15] is used for estimating the memory nonlinearity components. The time domain signal r(n) after PD is expressed as: r (n)  h0 

N1 1

N 2 1 N 2 1

 h (i ) x (n  i )    h (i , j ) x (n  i )x (n  j )  i 0

1

i 0

j i

2

N 3 1 N 3 1 N 3 1

   h (i, j, k ) x(n  i )x(n  j)x(n  k )..... i 0

j i k  j

3

, (1)

where h0 is the additive Gaussian noise brought by the device background noise and channel impairments [14], hK(…) (K>0) is the Kth order Volterra kernel corresponding to the linear (K=1) or nonlinear (K>1) responses. NK is the memory length of Kth order

i=1 Y(N)

Nonlinearity-aware Equalization

y’(n)

i>1 Decision & IFFT

Nonlinearity Estimation

Fig. 2. DSP flow of TD-NE.


(3)

P (n)   {P (n)  k (n)x (n) P (n)} ,

(4)

hˆ i (n)  hˆ i -1 (n)  e i (n)k i (n) .

(5)

1

i

i -1

i

i

T

i -1

̂ is updated by Eq. 5 with the In the iteration, the estimated 𝒉 error signal e(i) in Eq. 2 and the gain vector ki from Eq. 3 and Eq. 4. In Eqs. (2)-(5), the index n denotes the nth subcarrier, λ is a predefined constant before initialization, which was set as 0.9999 in our experiment. P(n) is the inverse correlation matrix of the input signal. After obtaining the nonlinear kernels with RLS algorithm and training symbols, the signal is equalized by subtracting the nonlinear noise with Eq. 6: ith

N 2 1 N 2 1

y '(n)  y (n)   i 0

 hˆ (i, j) x(n    i )x(n    j)  j i

2

N 3 1 N 3 1 N 3 1

   hˆ (i, j, k ) x(n    i )x(n    j)x(n    k ) , i 0

j i k  j

3

(6)

in which a delay factor α (i.e. the delay of the TD-NE equalizer) is inserted to change the position of reference taps. After equalization with Eq. 6, y’(n) is utilized as the input for DF module for the next iteration until the loop comes to the end. In the experiment, the length of IFFT and cyclic prefix are set to 1024 and 32, respectively. The corresponding subcarrier spacing is 89.84-MHz. For simplicity in the DSP implementation, the memory length N2 and N3 of TD-NE are kept the same. The output amplitude of the DAC is set to 450-mV after optimization and the measured central frequency under above driving condition is ~1549.8-nm. The modulated optical signal from modulator (~0dBm) is fed into a spool of SSMF of 0.8/1.6-km length. Because of the lack of trans-impedance amplifier (TIA) in the experiments, one erbium doped fiber amplifier (EDFA) together with a variable optical attenuator (VOA) is used before PD to study the receiver sensitivity. In the experiments, the input optical power of the PD ranges from -2 to 7-dBm. Our noise analysis according to [16] demonstrates that at higher received optical power, a system with an EDFA as a preamplifier has very similar total noise figures as

20,5 20,0

4

w/o TD-NE 1st iteration 2nd iteration 3rd iteration 4th iteration

2

19,5

1

19,0

0 4

18,5 18,0

(a)

1 2 3 4 5 6 7 8 9 10 11 12 13

Memory Length

21 50 100 150 200 250 300 350 400 450 Frequency (GHz)

4.5

1.6km 0.8km B2B

24 10 22

12

20

SNR (dB)

9.0 13.5 18.0 22.5 27.0 31.5 36.0 40.5

14

18

16

16

18

19 18

B2B 0.8km 1.6km

20

12

22 10 0

50

Subcarrier Index

Fig. 3. Probed channel SNR without TD-NE.

Traning Symbol =3 Traning Symbol =5 Traning Symbol =7 Traning Symbol =10 Traning Symbol =15 Traning Symbol =20

16 15

(b)

100 150 200 250 300 350 400 450

SNR before TD-NE = 18dB

17

14

14

24 8

Mean SNR (dB)

20 0 8 0

1.6km

3

-2 -1 0 1 2 3 4 5 6 7 8 9

Delay

SNR Gain (dB)

k i (n)  P i 1 (n)x i (n){  x i (n)T P i -1 (n)x i ( n)}1 ,

the one integrating a TIA with sufficient gain and bandwidth. Therefore, it is expected that in real implementation the EDFA and VOA (effective gain < 7.5-dB) can be replaced by the TIA, particularly when a certain system power margin is reserved to compensate performance degradation if any. The signal-to-noise ratio (SNR) at the receiver is firstly probed by sending 16-QAM modulated DMT signal with 430 subcarriers (i.e., electrical signal BW is ~38.63-GHz), which is shown in Fig. 3. The fast roll-off of SNR in the 1.6-km case is observed because of the power fading from the chromatic dispersion. The number of the used subcarriers is optimized to balance the mitigation of error propagation in TD-NE and overall system throughput. The optimized numbers of the used subcarriers are 430, 430 and 376 for the optical B2B, 0.8-km and 1.6-km fiber transmission cases, respectively. To investigate the effectiveness of TD-NE in improving the SNR, we measure the mean SNR under different memory lengths and different times of iterations in the optical back-to-back scenario, as shown in Fig. 4 (a). The result of conventional linear equalizer is also listed for comparison. The number of training symbols and the delay factor α are 10 and 2, respectively. It is clear that the mean SNR is improved with the increment of memory length of TD-NE. The SNR gain by TD-NE over LE gets saturated when memory length is larger than 8 regardless of the times of iterations. Besides, the SNR is also improved with the process of TD-NE iterations. There is a large improvement of SNR with the first and second iterations, and the additional SNR gain becomes small when the iteration number is larger than 3. It is because the RLS algorithm has converged to the optimal achievable solution. With 4 iterations and a memory length of 8, a SNR gain of ~2.2-dB is achieved by TD-NE. The impacts of the number of training symbols and the value of delay factor α on the performance of TD-NE are also investigated, which is shown in Fig. 4 (b). The memory length is 8 and the iteration number is 4. A fairly good performance is observed when the number of training symbols exceeds 10. The optimal range of α is 1~5 regardless of the actual number of training symbols. After optimization of these parameters, the SNR are significantly improved with TD-NE in all three systems with different transmission distances. Fig. 4 (c) shows that the SNR gains of different subcarrier channels varying from 0-dB to 3.5-dB.

Mean SNR (dB)

effect, x(n) is the transmitted electrical DMT signal. The DSP flow of TD-NE is shown in Fig. 2. The signal is firstly equalized by a frequency domain one-tap linear equalizer. The output from equalizer is denoted as Y(N) and its corresponding time domain sample is y(n). At the first iteration, Y(N) is fed to a decisionfeedback (DF) module followed by an inverse fast Fourier transform (IFFT) module to get the estimation of the transmitted temporal samples x(n), which is denoted as 𝑥̃(𝑛). The nonlinear kernels are estimated by comparing y(n) and 𝑥̃(𝑛). The estimation process is realized by recursive least square (RLS) algorithm with the training symbols as follows: (2) e (i )  y i (n)  x i (n)T hˆ i -1 (n) ,

0.8km

3 2 1 0 4 3 2

B2B

1 0 0 50 100 150 200 250 300 350 400 450

(c)

SC nr.

Fig. 4. (a) Mean SNR versus memory length with iterations, (b) mean SNR versus delay with training symbols, (c) SNR gain per subcarrier in the cases with 1.6-km SSMF, 0.8-km SSMF and optical B2B.

It’s a preprinted version for researchgate, all rights reserved by Optics Letters, OSA. B2B w/ TD-NE B2B w/o TD-NE

0.8km w/ TD-NE 0.8km w/o TD-NE


BER

1E-01

20% OH SD-FEC limit 2.7e-2

1E-02

-3 -2 -1 0 1 2 3 4 5 6 Received Optical Power (dBm)

7

8

Fig. 5. BER performance without bit-power loading. Regarding the transmission demonstration, the DMT transmission with 64-QAM is firstly carrier out. The measured BER in terms of the received optical power at PD is shown in Fig. 5. For all cases, with TD-NE, the BER is lower than the threshold of the SD-FEC with 20% overhead [17]. The achieved net rates are 136.7Gbit/s, 131.6-Gbit/s and 121.7-Gbit/s for optical B2B, 0.8-km and 1.6-km fiber transmission with power budget margins about 3-dB, achieving effective electrical spectrum efficiencies of 3.54-, 3.41and 3.15-bits/s/Hz, respectively. However, since the channel is a colored-SNR channel, and the BER is higher at high subcarrier frequencies. Thus, adaptive bits and power loading is used to further enhance the system performance and increase the system capacity. The bits and power loading profiles of 1.6-km transmission based on the probed SNR are shown in Fig. 6, and the received QAM constellations at B2B case are shown in the insets of Fig. 6. The shadow shows the received spectrum of DMT signal after detection. The gross date rates after bits and power loading are 215.4-Gbit/s, 209-Gbit/s and 200-Gbit/s for the three tested cases, respectively.

2,5 2,0

6

1,5 4 1,0 2 0

Power (a.u.)

Number of Bits

8

(a)

(b)

0,5 0,0 50 100 150 200 250 300 350 400 Subcarreir Index

0

(c)

Fig. 6. Bits and power loading profiles at 1.6-km SSMF transmission. 200 B2B w/ TD-NE B2B w/o TD-NE


w/o bit-power loading w/ bit-power loading


180

0.01

(a)

Net Data Rate (Gbit/s)

BER

0.1

20% OH SD-FEC limit 2.7e-2

-3

-2

-1 0 1 2 3 4 5 6 7 Received Optical Power (dBm)

8

179.5

174.2 166.7

160

140 136.7

131.6 121.7

120

100

(b)

0.0

0.8

Fiber Length (km)

1.6

Fig. 7. (a) BER performance with bit-power loading, (b) net data rates versus transmission distance.

The BER of DMT transmission after bit-power loading is shown in Fig. 7 (a). The achieved net rates are 179.5-Gbit/s, 174.2-Gbit/s and 166.7-Gbit/s for optical B2B, 0.8-km and 1.6-km fiber transmission with power budget margins larger than 2.5-dB. Effective electrical spectrum efficiencies of 4.65-, 4.51- and 4.93bits/s/Hz have been achieved, respectively. For the case with 1.6km SSMF, the net rate of DMT system with bit-power loading is improved by 37% compared to the 64-QAM DMT scheme and the spectrum efficiency is increased by 1.78-bits/s/Hz. The net data rates versus transmission distance is shown in Fig.7 (b). To conclude, by using a simple IM/DD system based on a packaged 1.55-μm DFB-TWEAM, a single DAC and a single PD, we experimentally demonstrated DMT transmission beyond 200Gbit/s (net rate beyond 166.7-Gbit/s) for high-capacity shortreach link. With nonlinearity-aware DSP, the measured BER is below the threshold of SD-FEC up to a 1.6-km SSMF link. Acknowledgments. This work was supported by Swedish Research Council, Swedish Foundation for Strategic Research, Göran Gustafsson Stiftelse, China Scholarship Council, and National Natural Science Foundation of China (#61605047, 61671212, 61550110240), and Natural Science Foundation of Guangdong Province (2016A030313438). The equipment was funded by Knut and Alice Wallenberg foundation. The authors thank Tektronix and Keysight for the loan of real-time oscilloscope and AWG, respectively.

References 1. Cisco Global Cloud Index: Forecast and Methodology, 2015–2020 2. IEEE P802.3bs 200 Gb/s and 400 Gb/s Ethernet Task Force 3. IEEE P802.3ba 40Gb/s and 100Gb/s Ethernet Task Force 4. S. Kanazawa, H. Yamazaki, Y. Nakanishi, T. Fujisawa, K. Takahata, Y. Ueda, W. Kobayashi, Y. Muramoto, H. Ishii, and H. Sanjoh, in Proceedings of Optical Fiber Communication Conference (OFC) (2016), paper Th5B.3. 5. H. Yamazaki, M. Nagatani, F. Hamaoka, S. Kanazawa, H Nosaka, T. Hashimoto, Y. Miyamoto, in Proceedings of European Conference on Optical Communication (ECOC) (2016), paper Th.3.B.4. 6. H. Mardoyan, M. A. Mestre, J. Manuel Estaran´, F. Jorge, F. Blache, P. Angelini, A. Konczykowska, M. Riet, V. Nodjiadjim, J. Y. Dupuy, and S. Bigo, J. Lightwave Technol., 35, 1253 (2017). 7. S. Lange, S. Wolf, J. Lutz, L. Altenhain, R. Schmid, R. Kaiser, C. Koos, S. Randel, and M. Schell, in Proceedings of Optical Fiber Communication Conference (OFC) (2017), paper Th5C.5. 8. M. Chagnon, S. Lessard, D. V. Plant, Photon. Technol. Lett., 28, 2233 (2016). 9. F. Li, J. Yu, Z. Cao, J. Zhang, M. Chen, X. Li, J. Lightwave Technol., 35, 941 (2017). 10. Y. Zhu, X. Ruan, K. Zou, and F. Zhang, J. Lightwave Technol., (2017). 11. H. Isono, T. Takahara, H. Sakamoto, Y. Miyaki, T.Tanaka, M. Nishihara, and J. C. Rasmussen, IEEE802.3bs 400GbE Task Force, Plenary Meeting 12. O. Ozolins, X. Pang, M. Olmedo, A. Kakkar, A.Udalcovs, S. Gaiarin, J. R. Navarro, K. M. Engenhardt, T. Asyngier, R. Schatz, J. Li, F. Nordwall, U. Westergren, D. Zibar, S. Popov, and G. Jacobsen, J. Lightwave Technol., 35, 1174 (2017). 13. P. S. Chow, J. M. Cioffi, and J. A.C. Bingham, IEEE Trans. Commun., 43, 773 (1995). 14. L. Zhang, S. Xiao, M. Bi, L. Liu, Optics. Commun.,364, 129 (2016). 15. X. Hong, O. Ozolins, C. Guo, X. Pang, J. Zhang, J. R. Navarro, A. Kakkar, R. Schatz, U. Westergren, G. Jacobsen, S. Popov, and J. Chen, in Proceedings of Optical Fiber Communication Conference (OFC) (2017), paper Th3D.3. 16. G.P. Agrawal, Fiber-Optic Communication Systems (2002). 17. D. Chang, F. Yu, Z. Xiao, N. Stojanovic, F. N. Hauske, Y. Cai, C. Xie, L. Li, X. Xu, and Q. Xiong, in Proceedings of Optical Fiber Communication Conference (OFC) (2012), paper OW1H.4.