1. Introduction

Owing to the explosion of Internet services, bandwidth requirement from end-users are growing rapidly. In 2015, IEEE 802.3 Working Group started the standardization process for next generation Ethernet passive optical network (NG-EPON). The objective of NG-EPON is providing 25-Gb/s capacity with one wavelength or 100-Gb/s capacity with 4 stacked wavelengths [1]. Also, 25-Gb/s per wavelength transmission is considered as an upgrade path for next generation PON stage 2 (NG-PON2). For cost control, 10-G class optics enabled 25-Gb/s transmission is preferred. The most common solution is using bandwidth-efficient modulation formats, such as electrical duobinary (EDB) [2, 3] and 4-ary pulse-amplitude modulation (PAM-4) [4, 5]. Another promising solution is using equalization techniques to compensate for the bandwidth limitation. We have reported a real-time demonstration of 4 × 25G time and wavelength division multiplexing (TWDM)-PON based on 10G directly modulated laser (DML) and 10G avalanche photodiode (APD) receiver [6]. The equalization is realized in optical domain by using a delay interferometer (DI) for spectral reshaping. On the other side, electrical and digital equalizers, which are widely used for Inter Symbol Interference (ISI) elimination, such as feed-forward equalization (FFE), decision-feedback equalization (DFE), and Maximum likelihood sequence estimation (MLSE) etc., can also be used to compensate the bandwidth limitation to enable 2-level detection [7, 8]. A 25-Gb/s Faster-than-Nyquist non-return-to-zero (FTN-NRZ) downstream transmission based on 10G-class optics is demonstrated by Jiangwei Man et al. Least mean square (LMS)-FFE, post-filter and MLSE are used successively to achieve −26 dBm sensitivity after 20-km standard single mode fiber (SSMF) transmission [9]. Minghui Tao et al. proposed to use narrow-filter-compensation (NFC), enhanced forward error correction (FEC) and MLSE to equalize the duobinary signal back to NRZ waveform [10]. In order to get a good performance, expensive real-time oscilloscopes with high sampling rate and high resolution are used in most of the previously reported demonstrations. But for the cost-sensitive PON applications, besides of the performance, the complexity of the signal processing and the cost of the devices should also be taken into consideration.

We have made an investigation on the requirements on ADC parameters for enabling MLSE detection in 10G-class optics based 100G-PON system, where the influence of timing jitter, resolution etc. on the sensitivity is analyzed in O-band [11]. In this paper, we made a comparison between the FFE/DFE based equalizer and the MLSE based digital detector. 10-G Electro-absorption Modulated Laser- Transmitter Optical Subassembly (EML-TOSA) and 10-G APD-Receiver Optical Subassembly (ROSA) are used as transmitter and receiver respectively. The experiment is carried out in C-band, where the chromatic dispersion introduced signal distortion is also considered. The equalizer is employed at the receiver side to compensate for the ISI from the bandwidth limitation and chromatic dispersion, and convert the signal from duobinary format back to NRZ-on-off-keying (OOK) format for 2-level detection. To investigate the cost introduced by the analog to digital converter (ADC) in MLSE solution, we made an evaluation of the requirements on ADC parameters for enabling MLSE detection, including the sampling rate, the timing offset, timing jitter and resolution, etc. Experimental results show that using a 25-GS/s ADC with 4-bit resolution, 25-Gb/s transmission is realized using 10-G TOSA and ROSA, and 28-/30-dB loss budget can be achieved in C-/O-band respectively. For achieving similar performances with MLSE detection, 13-tap FFE combined with 3-tap DFE are required.

2. Experimental setup

Figure 1 shows experimental setup. Both the transmitter and receiver are designed for 10 Gb/s application. For 10-Gb/s data modulation, the extinction ratio is 8.2 dB under 2V driving voltage. The APD-ROSA at the receiver side is biased at 26V, and the measured sensitivity at BER of 1 × 10⁻⁹ is ~-26 dBm for the EML-modulated 10-Gb/s signal in back to back (BtB) case. The transmission penalty is ~2-dB with 800 ps/nm dispersion. The frequency response of the transceivers is shown as inset (d) in Fig. 1. The pseudo random binary sequence (PRBS) data with a word length of 2¹⁵-1 is generated by a pulse pattern generator (PPG-Anritsu MP1800A) at 25 Gb/s data rate in NRZ-OOK format. Then the signal is directly modulated to the 10G EML-TOSA in optical line terminal (OLT). Inset (e) shows the optical spectrum of the modulated signal. After 25-km SMF transmission, the signal is launched into a variable optical attenuator (VOA) for power variation and sensitivity measurement. Inset (e) depicts the optical spectrum of the modulated 25-Gb/s signal. At the receiving side, the signal is sampled by a real-time oscilloscope (OSC) with 80-GS/s sampling rate and 25-GHz bandwidth. The nominal resolution of the OSC is 8-bit, and the effective number of bits (ENOB) is claimed to be 5.8. Then equalization processes, including FFE/DFE and MLSE are all performed off-line using Matlab. It can be seen from the inset that the signal is converted into duobinary format due to the bandwidth limitation of the transceivers. Besides, as the EML is operated in 1534 nm, the received signal is distorted by the chromatic dispersion after 25-km SMF transmission as inset (c) shows. Therefore, the off-line DSP module should be responsible for the ISI caused by both the bandwidth limitation and the chromatic dispersion.

 

Fig. 1 Experimental setup for 10G-class optics based 25Gb/s NRZ transmission system. Insets: (a) Electrical eye diagram before 10G EML; (b) Electrical eye diagram received by 10G APD for BtB. (c) Electrical eye diagram received by 10G APD for 25-km fiber transmission.

Download Full Size | PDF

The equalization process of FFE&DFE and MLSE are shown in Fig. 1. As the sampling rate of the OSC is 80 GS/s, the data should be resampled to an integer multiple of the data rate first, and then timing recovery, down-sampling, signal equalization, signal detection and BER calculation are performed. For the MLSE evaluation, an extra signal re-quantization module is added for evaluating the ADC resolution requirement. In our experiment, re-quantization is done after down-sampling to reduce the computing complexity. The simulation about the timing offset and timing jitter are performed during the down-sampling process by choosing data sampled on different time. After equalization, the distorted duobinary signal is recovered back to OOK format. Finally, after the NRZ decoding, the BERs of the signal sampled at different received optical powers are calculated for sensitivity evaluation. Note that the equalization is performed after down-sampling, and only one sample per bit is required for signal equalization. In real application, as long as a CDR is employed to find the optimal sampling time, 25GS/s sampling rate is sufficient for enabling signal equalization. Therefore, some of the DSP modules depicted in Fig. 1, including the up- and down-sampling module and the re-quantization module are all added for performance evaluation, which is not necessary in real-applications.

3. Experimental results

3.1 Evaluation of FFE&DFE based equalizer

Figure 2 shows the equalization performance of FFE and DFE filter with different number of taps in BtB case. In the experiment, the signal is sampled by the OSC under different received power varying from −18 dBm to −28 dBm with an increasing step of 1 dB. Then we calculate the BERs under different received power cases for each equalizing situation, thereby each situation would include several individual BER points. Then a BER curve as a function of received power can be obtained by linear fitting of these points. We take 1-M bits for BER calculation and define the received power at BER of 1 × 10⁻³ as the sensitivity in the experiment. The sensitivities referred in this paper are all calculated in this way. Without equalization, we cannot obtain a BER under the FEC limit. 5-tap FFE gives a sensitivity of −19.34 dBm at the FEC limit, and the sensitivity improves with the increase of tap number as shown in Fig. 2(a). −23.63 dBm sensitivity is achieved with 11-tap FFE, and it makes negligible improvement when we further increase the number of the tap. However, when we add DFE filter into the equalizer, the performance is significantly improved. With only 1-tap DFE added, the sensitivity can be increased by 2~3 dB. The performance of the combined DFE and FFE equalizer is depicted in Fig. 2(b). 5-state Viterbi demodulator gives a sensitivity of ~-25.8 dBm, which is depicted as the horizontal line for comparison. With 13-tap FFE and 3-tap DFE, −25.74 dBm sensitivity is achieved, which is comparable with the MLSE demodulator. When we increase the FFE taps from 13 to 15 and increase the DFE taps from 3 to 5 or even to 7, the sensitivity improvement is less than 0.05 dB, which is almost negligible as shown in Fig. 2(b). And when we further increase the taps beyond 15/7 for FFE/DFE, there is almost no further sensitivity improvement.

 

Fig. 2 Receiver sensitivity as a function of FFE and DFE filter taps at BER = 1 × 10-3 (a) FFE only vs FFE + DFE-1 in BtB case; (b) Different taps of DFE and FFE in BtB case; (c) Different taps of DFE and FFE after 25-km SMF transmission;

Download Full Size | PDF

The results for 25-km SMF transmission case are shown in Fig. 2(c). We set the tap number of the DFE filter at 1, 3, 5, 7, and ranges the FFE tap number from 3 to 15 to evaluate the equalization performance. The FEC limit cannot be achieved using only FFE, and at least 1-tap DFE is required to get a BER lower than 1 × 10⁻³. For DFE filter with 3-, 5- and 7-tap cases, the performances are quite similar, which are ~1-2 dB better than 1-tap case. The sensitivity improves with the increase of FFE taps until the FFE tap number exceeds 13. As a result, −22.83 dBm sensitivity is achieved by using 13-tap FFE and 3-tap DFE, which is ~3 dB worse than the BtB case. Further increase the tap numbers also makes negligible improvement. The performance of the MLSE demodulator is ~1.2 dB better than the FFE and DFE based equalizer, as Fig. 2(c) depicts. Figure 3 plots the eye diagrams of the received and the equalized signal, from which we can see the equalization effect clearly.

 

Fig. 3 (a) Eye diagrams of the received and equalized signal in BtB case and (b) 25 km transmission case.

Download Full Size | PDF

According to the results above, 13-tap FFE and 3-tap DFE filter provides comparable performance with MLSE detector in BtB case, and the sensitivity is still ~1 dB worse than MLSE in 25-km SMF transmission case. As for the complexity, 2-state MLSE requires 32 multiplication operations, which is twice as much as the FFE and DFE solution. And the computing complexity of MLSE increases exponentially with the increase of state numbers. As a result, MLSE detector provides a better performance than FFE&DFE based equalizer at the cost of a much higher computing complexity. Therefore, FFE&DFE based solution is more cost-efficient for PON application, while MLSE would show its advantage for the scenarios where the loss budget is extremely limited. On the other hand, the OSC we used in the experiment has 80GS/s sampling rate and 8-bit resolution, which is much too expensive for practical application. In the following section, we will investigate the impact of the ADC parameters on the system performance, including sampling rate, resolution, timing offset and timing jitter, etc. Also, for reducing the computation complexity, the performance of MLSE detector with different state numbers are also evaluated.

3.2 Evaluation on MLSE based detector

The computing complexity of MLSE detector is closely related with the number of states. Plenty of work has been done to reduce the computing complexity of MLSE algorithm [12, 13]. To verify the computing complexity of MLSE detector in our system, the performances of MLSE detector under different state number cases are evaluated first. We varies the state number from 2 to 5 and calculates the BERs, and the results are shown in Fig. 4(a). In BtB case, similar sensitivities are obtained when the state number varies from 2 to 5. 25-km SMF transmission introduces 2-dB sensitivity penalty for 3-, 4- and 5-state MLSE detection, and 1-dB more degradation is observed for the 2-state case. So the state number is set at 3 in the following evaluations.

 

Fig. 4 BER curves as a function of (a) MLSE state number (b) timing offset (c) timing jitter and (d) ADC resolution

Download Full Size | PDF

Then we evaluate the impact of ADC parameters on the signal sensitivity. Note that the channel response for the downstream link is relatively stable within a period of time. But for upstream link, adaptive channel estimation is required to determine the channel response of the bursts from different ONUs. Recursive least squares (RLS) algorithm is used in the experiment for channel estimation, where the channel response can be obtained within 150 training bits. And a time recovery module is necessary for both upstream and downstream links. Firstly, we evaluate the influence of timing errors coming from the clock recovery module or the timing jitter of ADC on the receiving sensitivity. The 80-GSa/s sampled data is first up-sampled to 800-GSa/s, corresponding to 32 samples per bit. Then absolute-law based symbol timing recovery (AL-STR) algorithm [14] is applied to find the optimal sampling time, i.e. the peak of the pulse. Estimation of the channel response is made based on the data obtained on the optimal sampling time. As the data is up-sampled by 32 times, the sampled data is reshaped into 32 columns, corresponding to the data sampled on 32 different timing offsets. Considering the data rate of 25 Gb/s, the time difference between two neighboring sampling points is 1.25 ps. Figure 4(b) shows the calculated BER in different timing offset cases using the data sampled on −22 dBm and −24 dBm as examples. The BER grows almost linearly with the timing offset increases, which verifies the necessity of the CDR module. As for the timing jitter, a random index varies within [-1, + 1], [-2, + 2]……or [-8, + 8] following Gaussian distribution is added to the column number during the down-sampling process to simulate the timing jitter of ± 1.25 ps, ± 2.5 ps…… or ± 10 ps cases. The calculated BERs are depicted in Fig. 4(c), where the BER increases gradually with the increase of timing jitter. With ± 10 ps jitter, the BER increases from 4 × 10⁻⁴ to 4 × 10⁻³ for −22-dBm received power. The results for the ADC resolution evaluation are shown in Fig. 4(d). The sampled data are re-quantified into 3-, 4-, …… 7-bit, and the BERs are calculated respectively. Similar performances are obtained when the resolution is decreased from 8- to 6-bit, verifying that the effective number of bits is ~6 as the manufacture claimed. During the off-line re-quantization, all the bits are occupied for describing data. Therefore, when we change the sampling resolution to 3-, 4-, and 5-bit, the ENOB is also 3-, 4- and 5-bit. And when we re-quantize the sampled data to 6-, 7- and 8-bit, the ENOB would be limited by the OSC to ~6-bit. When we further reduce the resolution, the BER increases rapidly. For 1- and 2-bit case, we cannot obtain a BER lower than 1 × 10⁻³, and a much higher received power is required to reach the FEC limit. Therefore, for loss budget consideration, 3-bit or higher resolution is required in the proposed system.

Finally, the relation between the system performance and the ADC parameters are evaluated. The ADC resolution is varied from 3- to 8-bit, and the timing jitter is varied from 0 to ± 10 ps with a step of 1.25 ps. The sensitivity of the received signal after 25km fiber transmission is calculated for all the cases and the results are listed in Table 1. For the region where the timing jitter is lower than ± 10 ps, and the resolution is higher than 4-bit, −23 dBm sensitivity can always be obtained. For the 4-bit case, the timing jitter should be limited to ± 3.75 ps to get a sensitivity lower than −23 dBm. Considering the 5-dBm output power from the EML-TOSA, 28-dB loss budget can be achieved using a 4-bit ADC with 25GSa/s sampling rate. 25-km SMF transmission in C-band can be supported. Further increase of the transmission distance introduces higher loss and severer signal distortion. Considering the limited loss budget in our system, it requires optical amplifier and more complex DSP to support a longer distance transmission, which is not demonstrated in this work. Besides, we have made an investigation in O-band using 25G-EML and 10G-APD. As the 2-dB sensitivity penalty introduced by the chromatic dispersion during fiber transmission is avoided, 30-dB loss budget can be achieved [11].

Table 1. Sensitivity at BER of 1 × 10⁻³ vs. resolution and timing jitter of ADC

View Table

4. Conclusion

In this paper, we investigate the performance of FFE&DFE and MLSE based equalization techniques for NG-EPON application. As 10-G transceivers are used for 25-Gb/s NRZ modulation, the equalizers are employed to compensate for the ISI caused by both the bandwidth limitation and chromatic dispersion during fiber transmission. Evaluations of the requirement on ADC parameters for enabling effective MLSE detection are made, including the sampling rate, the resolution, the timing offset and timing jitter. Experimental results show that 28-dB loss budget can be achieved using a 4-bit ADC with 25GSa/s sampling rate in C-band, which could support 25-km SMF transmission and at least 64 users. By comparison, 13-tap FFE and 3-tap DFE are required to get similar performances with MLSE based detector.