Open Access
Add to BibSonomyAbstract
112Gbit/sec DSP-based single channel transmission of PAM4 at 56Gbaud over 15GHz of effective analog bandwidth is experimentally demonstrated. The DSP enables use of mature 25G optoelectronics for 2-10km datacenter intra-connections, and 8Tbit/sec over 80km interconnections between data centers.
© 2015 Optical Society of America
1. Introduction
The dramatic growth in data centers capacity requirements results in newly evolved architectures and topologies such as spline-leaf and east–west configurations. In turn, a huge amount of interconnection capacity is required for data transmission within the data centers elements, and externally, for interconnection between data centers. Basic data center structure and interconnection architecture is depicted in Fig. 1. While various inter-and intra-connection links include top of the rack-, aggregation-, and core-switches, all the connections require the highest available transmission rates, i.e., 100Gbit/Sec and 400Gbit/Sec. The demand for this high throughputs and yet extremely low cost and low power data transmission raises new challenges to the optical communication technology, and particularly to the physical layer. Recently, increasing amount of research attention is focused on 100G transmission over single lambda based on 56Gbaus PAM4 modulation [1,2]. However, most published works are based on relatively wide bandwidth components of 35GHz and above, which are still under development phase and relatively expensive. In this paper, a unique 100/400Gbit/Sec DSP based integrated circuit (IC) solution is presented, delivering ultra-high capacity connections between short reach switches of 500m-2km/10km, as well as 8Tbit/sec long reach 80km DWDM interconnections over effective analog bandwidth of 15GHz. The solution is targeted to work with mature and available optical components, a fact that reduces further the expected cost of the systems. The proposed solution incorporates PAM-4 modulation with advanced DSP technology that enables capacity optimization in a single lambda 100Gbit/Sec solution, the most flexible and manageable solution for switch-to-switch connectivity. This technology allows an elegant data center interconnection architecture, with basic 100Gbit/Sec link per lambda or a basic 400Gbit/Sec link per quad CWDM, with the usage of a duplex fiber. The solution reduces dramatically fiber infrastructure (which is today based on multiple fibers).
Fig. 1 (a) Data centers new architectures and interconnections, and (b) Interconnection between data centers; up to 8Tbit/sec with 80 DWDM channels over a single compensated fiber.
2. System model
The proposed single channel 100G solution is based on an optical transceiver module that includes a “clean” architecture of a single transmitter optical sub assembly (TOSA) and a single receiver optical sub assembly (ROSA) that interface to the DSP based electronic IC. Both TOSA and ROSA are based on mature and commercially available 25G components, including the laser, modulator, and driver on the transmit side, as well as the photo-detector and trans-impedance amplifier (TIA) at the receive side. Figure 2 presents the high-level block diagram of the electronic IC. It includes a standard bi-directional 100Gbit/sec electrical interface on the “client” side (4x25G), a main DSP engine, and an electrical analog interface on the “line” side, interconnecting with the optical elements. The latter includes a digital to analog converter (DAC) at the transmit path, as well as an analog-to-digital converter (ADC) at the receive path, both operating at 56Gsamples/sec, with 6-bit of effective resolution, and 3-dB analog bandwidth of less than18GHz.
3. DSP challenges and capabilities
The DSP engine and the PAM4 demapper particularly, are the heart of the solution, enabeling the transmission of PAM4 at 56Gbaud rate over less than 20GHz of analog bandwidth (10dB point). The extremely narrow bandwidth (BW) as well as the different impairments of the optical channel, requires aggressive ISI mittigation under challanging power and area limitations. The PAM4 demapper allows PAM4 decoding at pre-FEC BER values better than 1E-3. In turn, the PAM4 decoding procedure is further enhanced using a relatively light forward error correction (FEC), that yields “error-free” decoded data, i.e., better than 1E-15 BER. The main challenging DSP tasks are to compensate for the combination of the following distortions and noises: (1) strong intersymbol interference (ISI) associated with the narrow channel bandwidth (20GHz 10dB point vs. 56Gbaud transmission), (2) multi level symbols (PAM4), (3) non linear behavior of the optoelectronics components, e.g., exponential curve of the electro-absorption modulator laser (EML), or sine curve of the Mach-Zender Modulator (MZM), (4) residual chromatic dispersion of the single mode fiber, e.g., 2km fiber at 1550nm wavelength transmission, (5) multiplicative noise, e.g. relative intensity noise (RIN), (6) multi path interfercnce (MPI) due to reflections within the optical path, and (7) thermal noise. In addition, in the 80km interconnection case, where multiple WDM channels are transmitted, additional distortions and noise mechanisms are present, including (8) amlified spontaneous emission (ASE) noise which is non Gaussian and multiplicative, (9) additional residual dispersion resulting from second order chromatic dispersion slope, (10) fiber non linearities, (11) inter-channel crosstalk, and (12) additional ISI introduced by the 50GHz WDM multiplexers or interleavers.
It is well known from the early 1970’s [3] that the optimal solution for ISI mitigation is the maximum likelihood sequence estimator (MLSE). This detection scheme also includes inherent non-linear compensation capabilities, thus adequate for compensation for some of the non-linear channel impairments mentioned above. However, the full realization of such an equalizer, for the case of PAM4 vocabulary, and for the large channel memory depth of 8 taps, using the Viterbi algorithm (VA) will require more than 65,000 MLSE states, which is essentially not implementable.
Here, a novel approach is proposed [4], performing strong 15-taps feed forward filtering shortening the channel impulse response and also shaping the signal and noise. The resulting signal is processed by a reduced complexity MLSE (RC-MLSE) containing 256 PAM4 states, mitigating the ISI and compensating for the nonlinear distortions. This combined DSP solution is optimized in the sense of minimizing the overall distortion variance, thus maximizing the SNR at the PAM4 decoding stage. The outcome of the proposed DSP is a cost effective, powerful equalization and sequence detection scheme based on a combination of feed forward filtering and RC-MLSE, suitable for high order modulation under tough optical channel conditions.
4. Experimental results
An inclusive set of off-line experiments was performed, evaluating the performance of the proposed DSP approach, under broad range of conditions. In all cases, the ADC and DAC hardware described in section 2 was used. The experiments included several TOSA options, including 1310nm EML, 1550nm EML, and 1550nm MZM, all characterized with analog bandwidth in the 25GHz range. All the tested TOSAs were modulated at low extinction ratio (ER) values, between 4 and 6dB. The MZM was directly modulated by the DAC, as described in Fig. 6, while the EMLs were modulated using a differential to single ended driver with 1.5V driving voltage, yielding 5.5dB of extinction ratio, and 35dB of common mode rejection ratio. Several ROSAs were used with referred input noise at range ofup to . Fiber lengths of 2-10km were used for 1310nm tests. Fiber lengths of up to 5km (un-compensated), and up to 80km with dispersion compensating fiber (DCF) and optical amplifiers (similar to Fig. 1(b)) were used for 1550nm tests. In all the tests, the pre-FEC BER at the RC-MLSE output was better than 1E-3. Measured results of typical transmission over 10km of single mode fiber (SMF) at 1310nm are depicted in Fig. 3. The channel impulse response (CIR) and the resulting signal and noise spectra of the measured PAM4 56Gbaud signal at the receiver are presented in Figs. 3(a) and 3(b), respectively. Figure 3(a) reveals that there are 8 taps with non-negligible power in the CIR, indicating the sever ISI introduced by the narrow analog bandwidth of the cascaded optoelectronic components. Conversely, Fig. 3(b) reveals the measured spectrum of the received signal. Since the transmitted data is a random sequence with white spectrum, the measured received signal spectrum is essentially the frequency response of the overall channel, indicating a 3dB bandwidth lower than 15GHz.
Fig. 3 (a) Channel impulse response, 8 taps are observed; (b) measured received signal spectrum, and estimated noise spectrum, vs. normalized frequency.
In Fig. 4, two stages of the de-mapping process of a typical measurement of 100G over 2 km link are monitored and presented: Fig. 4(a) presents the raw recorded data immediately after the ADC, i.e., prior to any DSP action. Figure 4(b) presents the digital signal after the “mid-stage” feed forward filtering and interference cancellation. This is followed by the RC-MLSE PAM4 decoder and, in turn, the BER of the decoded PAM4 signal is monitored and reported. In Fig. 4(a) it is clear that the four levels of the received PAM4 signal can’t be identified, and their histogram looks similar to random Gaussian process. In Fig. 4(b) the four PAM4 levels can be identified after the “mid stage” DSP (each obeys Gaussian distribution). The SNR at this point however, is still poor, and “reasonable BER” cannot be obtained. The BER at the output of the DSP engine is reported to be better than 3E-4 in all the recorded time frames in this experiment. As a result, error free data stream at the FEC decoder output is expected.
The interesting application of long reach 80km transmission, using compensated and amplified 80km SMF, was also successfully tested and demonstrated. The experimental setup is presented in Fig. 5. The baud rate was increased to 60 Gbaud in order to allow 20% transmission overhead for potential usage of a stronger FEC. A full C-band WDM transmitter with external MZM is directly modulated by a 60GHz DAC, and the transmitted 60Gbaud PAM4 signal propagates through the 80km SMF including a two-stage EDFA and DCF. As the DAC electrical output is 1Vp-t-p differential, and the V-pi of the MZM is 3.5V differential, the extinction ration achieved by this driving scheme is only 4.5dB. In turn, the optical signal at the receive side is filtered by a 50GHz grid interleaver cascaded with a 100GHz WDM demux, attenuated by a variable optical attenuator, converted to electrical signal by the ROSA, and finally sampled and quantized by the 60GHz ADC. It should be emphasized again that all the optoelectronics are commercially available 25G components.
Figure 6 presents measured BER results versus received optical signal to noise ratio (OSNR), summarizing a set of offline test results. The OSNR was measured using an optical spectrum analyzer (OSA) within a 0.1nm band, immediately after the 100GHz WDM demux. In the figure, two BER curves are presented: the upper curve (blue triangles) includes measured results with a 42GHZ double-side bandwidth interleaver, while the lower curve (green squares) includes measured results for the same setup without the interleaver. The OSNR penalty associated with the use of the 42GHz interleaver is explained by the additional ISI imposed by the optical filtering effect. The 42GHz 3dB double-side-band optical filtering is equivalent to 21GHz 6dB single side band electrical filtering, which imposes additional ISI on the 60Gbaud PAM4 transmitted data. In addition, the graph contains information regarding the launch power into the fiber “LP”, and the input power into the DCF “DCF in” for each test. It should be noted that the interleaver introduces additional insertion loss of 2dB, therefore, in cases that the optical amplifiers are near saturation (at high OSNR values), the system with the interleaver yields slightly lower OSNR values as compared to the system without the interleaver, even though using same launch power. For all the cases of OSNR values higher than 40dB, the resulting BER values are better than 5E-3. This BER level of 5e-3 is the limit of the considered FEC, meaning that pre-FEC BER values below this threshold are expected to result with error free data after the FEC. For the case of no interleaver, the BER is better than 1E-3 for OSNR higher than 40.5dB. It should be mentioned that a third set of tests using a wider interleaver with double-side bandwidth of 46GHz was carried out, which yielded the same performance as the case of no interleaver. This observation implies that the proposed scheme allows throughput of 112 Gb/sec over 46GHz. Additional observation is that very good BER results were obtained under high launch power conditions, as high as 10.7dBm into the SMF. These low BER measurements under such high launch power include strong self-phase modulation (SPM), which is a signal-dependent deterministic distortion. Consequently, the fact that the BER results were kept very low proves the robustness of the MLSE scheme in covering and mitigating strong nonlinear distortions.
5. Comparison between experimental results and theoretical prediction
Inclusive comparison analysis was carried out between the measured BER results for the different test cases and known theoretical predictions. In Fig. 7, theoretical and measured BER curves are presented vs. the measured SNR at the output of the MLSE, noted as “SNR MLSE”. The black curve represents the theoretical-analytical match-filter bound in the case of an ISI-free system and additive white Gaussian noise (AWGN) [3]. The green curve represents the theoretical limit for a heavily induced ISI system due to bandwidth limitations plus AWGN, fully compensated by an optimal brute force MLSE receiver without any complexity limitations. This ISI-induced theoretical curve is calculated via Monte-Carlo simulations. The blue dots represent the entire set of BER measurements using the proposed DSP engine presented here. The SNR values span over a broad range of 6dB, between 16.5dB in the lowest case, representing the system with EML TOSA of lowest bandwidth (that introduces residual ISI) and highest non-linearity, up to SNR value of 22.3dB in the highest case representing the system with MZM TOSA at back-to-back fiber link. The red curve is a polynomial fit to the different BER measurements marked by the blue dots. It is clear from Fig. 7 that although the fact that the optical channel includes the special impairments described above, some very different than AWGN model, the results reported here are very close to the theoretical limit of a linear system with induced ISI and AWGN (green curve), up to a fraction of a dB of difference.
6. Conclusions
Single channel 100Gbit/sec transmission scheme over 15GHz effective analog bandwidth is presented based on PAM4 modulation at 56GHz baud together with DSP based de-mapping. The solution supports both 100/400Gbit/Sec short reach data center intra-connection (500m-10km) and 8Tbit/sec long reach 80km interconnections between data centers. Extensive set of experiments were performed, including wide variety of configurations and link distances, all yielding very good results that are proven to be very close to the theoretical limit, suggesting the robustness and superior performance of the solution.
References and links
1. X. Song, J. Man, and L. Zeng, “Investigation of 4x112Gbps PAM4 configuration for the km SMF PMD,” presented at the IEEE 802.3bs 400GbE task force (May 2014).
2. M. Mazzini, “Technical feasibility of 56Gbaud PAM4 optical link budget based on experimental measurements,” presented at the IEEE 802.3bs 400GbE task force (August 2014).
3. J. G. Proakis, Digital Communications (McGraw-Hill, 1995).
4. G. Dorman, D. Sadot, and A. Gorshtein, “Enhanced equalization based on a combination of reduced complexity MLSE and linear equalizer for heavily ISI-induced signals,” US Patent Provisional 61/946,960 (March 2014).
