1. INTRODUCTION

To greatly outperform its predecessors from the aspects of speed, latency, and reliability, fifth-generation (5G) mobile access has the ambition to support a broad range of use cases over a single mobile network. Fulfilling this ultimate goal is not only about revolutionizing wireless technology, but also disruptively transforming mobile network architecture and design. One challenge among many, which was often overlooked in the previous generations, is to develop a cost-effective optical transport network (OTN) capable of high bandwidth, low latency, and great scalability, from the core all the way to the edge and antenna sites. With the adoption of the centralized radio access network (C-RAN) in 4G/Long-Term Evolution (LTE), the baseband processing unit (BBU) can be separated from the remote radio head (RRH), offering resource pooling at a centralized location [1,2]. Typically, the term “backhaul” refers to the connectivity between the BBU and the mobile core, while “fronthaul” refers to the connectivity between the BBU and the RRH. While 4G backhaul is Ethernet-based today, fronthaul uses the Common Public Radio Interface (CPRI)—a protocol carrying time-domain digitized I-and-Q radio signals along with other management and synchronization data in a constant bit rate (CBR) time division multiplex (TDM) format. This can lead to a fronthaul data rate of tens or even hundreds of gigabits per second that is 1 or 2 orders of magnitude higher than the user data rate [3], which makes CPRI-based fronthaul impractical for 5G and beyond.

Therefore, the 3rd Generation Partnership Project (3GPP) identified eight possible functional splits (FSs) and introduced new functional blocks and interfaces since Release 14 for the 5G radio access network (RAN) architecture [4], also known as next-generation RAN (NG-RAN), consisting of the 5G core network (5GC) and 5G radio base station next-generation NodeB (gNB). The gNB is defined by three key functional blocks, namely, a central unit (CU), a distributed unit (DU), and a radio unit (RU), whereas in the 4G RAN only a single FS is specified between the central BBU and remote RRH, as depicted in Fig. 1. While the F1 interface defines the high layer split (HLS) between the CU and the DU, there are different options for the low layer split (LLS) between the DU and the RU. The HLS has similar throughput and latency requirements as the backhaul, whereas the LLS requirements are more stringent in both aspects. The trade-off of employing different LLS options in RAN deployments is not clear-cut, and there is no one-fit-for-all choice, but such a functional decomposition offers more flexibility to design and deploy the transport network.

 

Fig. 1. Evolution from 4G to new 5G RAN architecture.

Download Full Size | PDF

Ethernet is becoming the common data transport protocol regardless of the choice of a particular FS because of maturity, low cost, and ubiquitous use. Ethernet can enable meshed connectivity between RAN functional modules, statistical multiplexing, multivendor interoperability, and efficient wavelength resource use. As defined in the enhanced CPRI (eCPRI) transport requirement specification [5], a packet transport network based on Ethernet is adopted to carry the LLS data. The eCPRI specification defines a packet layer that provides specific services, mainly the user plane data, to the upper layers, while it inherits the existing Ethernet protocols for operations, administration, and maintenance (OAM); control and monitoring [C&M, like Simple Network Management Protocol (SNMP)]; as well as synchronization [like Precision Time Protocol (PTP) and synchronous Ethernet (SyncE)]. This natively opens up a huge potential to leverage the well-established market scale of Ethernet products (e.g., aggregator, router, switch) for fronthaul transport.

In this paper, we first address major challenges in 5G fronthaul transport and candidate solutions in Section 2. We then report our recent research innovations to overcome these challenges. In particular, we propose a wavelength-agnostic wavelength division multiplexing (WDM) access technique (Section 3) and a high-speed gray interface employing advanced modulation schemes (Section 4) on the optical layer to increase fiber capacity and efficiency. On the packet layer, we implement the time-sensitive networking (TSN) based on IEEE 802.1CM to eliminate latency variation (Section 5). Finally, we summarize the deployment options for the 5G RAN.

2. CHALLENGES IN 5G FRONTHAUL TRANSPORT

A. Cost-Effective Capacity Growth

As long as the FS of the fronthaul is above the low-PHY layer (i.e., Option 1-7 [4]), the transport throughput scales with the actual user traffic on the air interface plus a certain amount of overhead that is dependent on the particular split. This is in contrast to CPRI, where the fronthaul traffic comprises digitized baseband signals that are independent of the air/interface use. eCPRI, while not specific about split preference, intended first and foremost to achieve a reduction in fronthaul traffic, by placing the fast Fourier transform (FFT)/inverse FFT (IFFT) function at the RU. This opens up the potential to transport frequency-domain IQ samples rather than time-domain samples. Furthermore, as functions such as resource element de-/mapping are also moved into the RU, the fronthaul traffic can be reduced by avoiding the need to transmit nonactive subcarrier information. Due to the variable bit rate traffic now realized, it makes sense to use a packet-based transport solution that enables statistical multiplexing. Besides, the latest O-RAN specification (in the Open Fronthaul Interfaces Workgroup) also aimed at throughput reduction and uses the same Ether-type fronthaul transport [6]. Assuming the adoption of eCPRI, it is foreseen that 10 to 25 Gbit/s optical interfaces will be adequate for 5G fronthaul in the near term [7]. At large cell sites with multiple RUs, it is feasible and beneficial to aggregate fronthaul traffic to 100 Gbit/s between cell site gateways and DU sites.

To date, the most advanced fiber transmission techniques are already breaking the capacity barrier of 100 Tbit/s per fiber link [8]. With the evolution of coherent optics, 400 Gbit/s and more per wavelength become a realistic future perspective also in access and aggregation networks. From the technical point of view, the capacity demand of the fronthaul could therefore be attained in one way or another. However, the associated cost is of concern in this competitive market segment. It is challenging to scale up the capacity while keeping the cost per transported bit low. One solution is therefore to leverage low-cost components and eliminate less critical components where possible. Fortunately, in most deployment scenarios, the fronthaul reach is relatively short (${\lt}{{20}}\;{\rm{km}}$) due to the maximum acceptable latency specified in the fronthaul protocol. Thus, the optical link budget and other specifications may be relaxed in order to save the cost. In addition, statistical multiplexing can relax the throughput requirements when using Ethernet aggregation.

B. Deterministic Latency

Latency becomes more critical in 5G transport than in the previous generations, especially in tactile and immersive use cases, as the latency largely determines the end-user experience. There are two latency aspects that need to be taken into account for 5G: the first is user latency of the order of millisecond performance, measured at the egress of the user plane function, and the second is latency in relation to RAN signaling processed by the DU such as hybrid automatic repeat request (HARQ). Delays over the fronthaul interface can lead to performance inefficiency on the RAN physical layer, and this leads to a stringent latency requirement in the order of 50 to 250 µs round-trip time (RTT) between the RU and the DU. The light propagation time in the fiber is approximately 5 µs/km, which dominates the end-to-end latency. In the case of Ethernet aggregation or switching, queuing and buffering introduce latency from a few microseconds to tens of microseconds. Similar latency contributions also arise when using TDM-based networks, such as passive optical networks (PONs) or OTNs. Since the forward error correction (FEC) will be largely inevitable for the interface at 25 Gbit/s and beyond, the latency of FEC processing also needs to be taken into account. In comparison, optical components (e.g., transceivers, splitters, filters) contribute only a negligible latency in the range of nanoseconds.

However, what is more important is the determinism of latency. The variation in the one-way delay (OWD) of packets is called “jitter” (or “delay variation”). This term is used in different ways by different groups, as described in IETF RFC 3393 [9]. Jitter commonly has two meanings: the first one refers to the variation of a signal with respect to a reference clock signal; the second one refers to variation in delay where a single reference is chosen. The latter is the definition of the packet delay variation (PDV) used for the two-point Ethernet PDV described in ITU-T Recommendation Y.1563. When the reference packet is the one with the minimum OWD, the PDV is positive (or zero). The shape of the PDV distribution is identical to the delay distribution but shifted by the reference delay.

Fronthaul based on a LLS transported in Ethernet is prone to PDV influence [10], since the delay incurred between packet reception on the Ethernet interface and delivery into the higher layer RAN L2 functions is variable as a result of aggregation queuing and buffering mechanisms. The PDV impact is further exacerbated, given that the same Ethernet interface needs to distribute timing information, such as SyncE [with associated Synchronization Status Messages (SSMs)] and PTP [11], for time division duplex (TDD)-based radio operation as well as new radio technologies such as carrier aggregation and coordinated multipoint transmission and reception.

C. Deployment and Operation

Unlike a single split (i.e., CPRI) between the 4G/LTE BBU and RRH, the 5G RAN introduces a disaggregated architecture, where the individual functional blocks CU, DU, and RU can be flexibly placed in the network according to the operator’s requirements and constraints. In the following discussion, we consider the deployment scenario of a collocated CU and DU, as the LLS necessitates more capacity and less latency. Currently, early-deployed 5G networks and more LTE deployments use direct fiber connectivity, meaning a dedicated fiber pair between the central location and the antenna site (usually per trisector antenna). Although such a deployment benefits from the lowest cost gray optics, it is only feasible in a fiber-rich area. However, fiber resource is at a premium and scarce in dense urban areas, and sometimes it is even impossible to deploy new links. Moreover, densification of cell sites as well as small cells demand more fronthaul connectivity. Different optical underlay technologies can be considered to save the fiber resource. WDM allocates dedicated wavelengths for each CU/DU and RU pair instead of a physical fiber pair, increasing the total capacity over a single fiber. PON technology, widely used for fiber-to-the-home (FTTH), is another alternative to multiplex fronthaul flows, although the major drawback is nonnegligible and sometimes asymmetric latency added by the PON, as remote optical network unit (ONU) sites are aggregated at the optical line terminal (OLT). Higher in the Open Systems Interconnection (OSI) stack, at the data link layer, the Ethernet protocol already supports aggregation of low-speed interfaces up to 400 Gbit/s. The OTN could also be used to enable TDM-based multiplexing, yet without statistical multiplexing gain.

The other practical aspect is the network operation on the RU side. Since the transport equipment is alongside the remote radio equipment commonly deployed in an outdoor environment, the optical transceiver has to be temperature-hardened for a large operating range. Due to distributed locations of cell sites, it is difficult to manage and maintain the equipment onsite.

In the following sections, our proposed solutions will address these above challenges.

3. WAVELENGTH-AGNOSTIC WDM ACCESS

Transceivers that are tunable over the full C-band have been available for some time for applications in metro and long-haul networks. As compared to fixed-wavelength transceivers, inventory sparing of these devices is alleviated, as fewer devices are required to cover all channels. However, a tunable transceiver usually comes at a significant cost premium over a comparable fixed-wavelength transceiver. One reason for this cost increase is the additionally required calibration of the module to tabulate all tuning parameters for each wavelength and over the full environmental temperature range. Furthermore, tight wavelength control to within ${\pm}{20}\;{\rm{pm}}$ requires a wavelength locker in each module. While WDM technology in 5G access networks would be a good solution to solve the capacity and latency requirements, the high transceiver cost would be prohibitive for large-scale deployment. Fixed-wavelength transceivers, on the other hand, would be operationally difficult to deploy in a fronthaul environment, as each port would need to be equipped with the correct transceiver and all wavelength types of transceivers would need to be stocked for repair.

Therefore, ITU-T SG15 developed a standard for tunable transceiver interfaces, where the wavelength of the remote unit is controlled by the CU. Equipment required for wavelength accuracy is located at the CU and shared between a number of remote units, also leading to cost-sharing. This standard, developed under the working title “G.metro” and released as Recommendation G.698.4 [12], specifies optical interface parameters as well as a communication channel between the CU and the remote units. In the following, details of the implementation are discussed.

A. Enabling Technologies

5G fronthaul systems generally connect a CU/DU with a number of RUs. As shown in Fig. 2, depending on the fiber infrastructure, the architecture of the system can be structured as a tree or drop-line. In the tree structure, all channels are sent via a trunk line to a branching point, which, for a WDM system, can contain a demultiplexing filter, like an arrayed waveguide grating (AWG). From here, each RU is connected by a separate fiber. In the drop-line structure, the trunk fiber passes each RU, and filters in optical add/drop multiplexers (OADMs) are used to drop the channels destined for a RU and add channels from the RU to the CU/DU. To reduce cost and to achieve symmetric propagation time, signals from a CU/DU to a RU and vice versa use the same fiber in different wavelength ranges.

 

Fig. 2. (a) Tree-structured fronthaul network with an AWG as a branching point, (b) drop-line fronthaul network with OADMs.

Download Full Size | PDF

While the routing of the signals between a CU/DU and a RU depends on their wavelengths and the filter ports to which the RUs are connected, no wavelength-specific transceivers are used. This requires the transceiver to tune its transmission wavelength according to the filter port. The tuning is based on a message from the CU/DU to the RU (on the respective CU/DU-to-RU wavelength) containing the operating wavelength of the RU. To reduce the calibration effort for the RU, the transceiver might not exactly know its tuning parameters to achieve the required wavelength. For this case, G.698.4 in its current version contains a start-up scheme, in which the RU transmitter sweeps its wavelength, at a reduced transmit power, until its signal is detected by the CU/DU. The power reduction is required to avoid cross talk from the tuning RU into already operating channels.

As a first step, a communication between the CU/DU and the RU must be set up before the RU starts its tuning procedure. As in most cases, the fiber is the only connection between the CU/DU and the RU; a message channel is defined in G.698.4. As this channel needs to operate for various protocols of the data channel, it was not embedded into the data channel, but rather implemented as an overhead channel by modulating the average power of the optical channel.

For the RU-to-CU/DU direction, a similar message channel was defined. At the same time, however, an identifier of the channel is required that allows the CU/DU to determine the presence of new channels [13] when tuning at a low power. The same identifier can also be used in the wavelength locker, common to all channels at the CU/DU, to distinguish the operating channels [14,15]. In G.698.4, pilot tones in the range of 50 kHz (${\pm}{2.5}\;{\rm{kHz}}$ with 50 Hz spacing) were chosen as an identifier. The tone frequencies were chosen high enough such that the gain of any optical amplifiers in the signal path would not follow the power modulation and generate cross-modulation and low enough such that the spectrum of the data signals transported over the wavelength would not be distorted by the pilot tone. The process is similar for the message channel data, which were directly modulated onto the optical signal. Manchester encoding [16] was chosen at a data rate of 50 kbit/s, such that the spectrum of the message channel would also be centered at 50 kHz. In Section 3.B, the performance of the data channels in the presence of the message channel and vice versa are discussed.

Several technologies for a low-cost implementation of a tunable transceiver have been developed in recent years. Two approaches use vertical cavity surface emitting lasers (VCSELs), with the tuning based on a micromechanical change of the cavity length [17,18]. An integration of the tunable laser with all optical receiver components (filter, photodetector) was demonstrated on a polymer platform [19]. Passive components, such as micro-optics, thin-film filters, and mirrors, as well as active elements such as gain blocks and photodiodes are assembled on a low-cost polymer waveguide chip.

Figure 3 shows a waveguide chip with assembled components as the target outcome of the PolyPhotonics project [20]. Other structures, e.g., polarization beam splitters and rotators, thin-film filters, or single-mode fibers can be added using the polymer platform [21–23]. The tunable Bragg laser is combined with an InP gain element, butt coupled to the waveguide chip, where the Bragg grating and a phase-shifter section are implemented as polymer waveguides. The laser can be directly modulated with data at 10 Gbit/s. The wavelength tuning mechanism works by controlling the temperature of the grating and the phase-shifter polymer. A high tuning efficiency is achieved by the high thermo-optical coefficient (TOC) of the polymer material (${\rm{TOC}} = - {{1}}{{{0}}^{{\rm{- 4}}}}/{\rm{K}}$), together with its low thermal conductivity. The TOC results in a frequency tuning coefficient of approximately 20 GHz/K, such that a temperature swing of 130 K of the Bragg section provides a tuning range of 2.5 THz, or 20 nm, sufficient for the half C-band tuning of G.698.4-compatible RU devices. This tuning range is limited at low temperatures by the operating temperature of the chip. Reducing the operating temperature would increase the tuning range, but also increase the power consumption required for cooling. At high temperatures, the tuning range is limited by the maximum allowed polymer waveguide temperature. The temperature of the Bragg grating and phase-shifter section waveguides must not exceed 200°C, limiting the tuning temperature swing to approximately 180 K.

 

Fig. 3. Waveguide chip with assembled components.

Download Full Size | PDF

To stabilize the set wavelength over time, the heating currents to the Bragg section and the phase-shifter section must be controlled. Temperature changes lead to a detuning of the lasing frequency. Therefore, the temperature of the laser must be kept constant, or the wavelength changes must be compensated for by changing the Bragg and phase currents. The first approach requires a mapping of the lasing wavelength at the operating temperature, varying Bragg and phase heating currents. For a stable operation, the temperature of the laser must be controlled, and especially cooled to guarantee the validity of the mapping. Additional hardware effort and power consumption are required for this approach. As an alternative, the wavelength is controlled either by a wavelength locker in the transceiver module or, as provided for in G.698.4, by a wavelength monitor in the CU and a feedback signal to the RU. For this approach, the mapping of the laser properties also includes varying the laser temperature. A relation between the Bragg and phase-tuning currents must be found such that adapting both avoids mode jumps [21].

Over time, stabilization is enabled by using a fixed ratio between Bragg and phase currents for temperature correction. Based on the feedback from the wavelength locker, the laser is tuned with a corresponding ratio until the desired wavelength is re-established. An accuracy of better than ${\pm}{1.25}\;{\rm{GHz}}$ for a period of more than seven days was achieved over a temperature range from 15°C to 60°C, sufficient for a 50 GHz channel spacing [21].

The tight integration of all components on a single optical chip allows a low-cost assembly of the transceiver device. While approaches with VCSELs and waveguide lasers can both use direct modulation up to a data rate of 10 Gbit/s, the chirp in waveguide lasers is reduced due to the longer laser cavity, allowing a longer transmission distance. This is especially noticed when a low-frequency pilot tone or a slow message channel is added to the modulation signal. While the frequency of a waveguide laser remains sufficiently constant, a VCSEL experiences a large wavelength modulation.

B. Performance Characteristics

The whole hardware has to be placed in an SFP+ package. As mentioned before, a message channel is needed for wavelength tuning at the remote site. A microcontroller (µC) in the SFP+ should be able to receive and decode the signals. Furthermore, it has to process the data and control the laser wavelength.

Conventionally, an out-of-band message channel is often implemented as a subcarrier, analog-modulated in amplitude or frequency. This requires additional analog hardware, like a digital-to-analog converter (DAC) and an analog-to-digital converter (ADC). The available µC must be able to perform the decoder functions, or additional programmable logic often is required.

To reduce the additional hardware effort, instead of an analog generated subcarrier, we used a direct digital modulation of the optical power by modulation of the laser bias current. Manchester coding is used as the modulation format, enabling DC balance and including the clock information in the coded signal. The µC is able to directly generate the digital message channel. The digital decoder can also be implemented in the µC. The message channel is extracted by a current mirror and then amplified by a logarithmic amplifier (Log-Amp), enabling a receiving sensitivity over eight decades. A digital low-pass filter (LPF) 5th order separates the message channel frequencies from the payload frequencies. The output from the LPF is directly connected to the µC. An implementation example is shown in Fig. 4.

 

Fig. 4. Receiver part with out-of-band communication (OOBC) electronics.

Download Full Size | PDF

The modulation depth of the Manchester envelope modulation needs to be carefully chosen such that the impact on the payload channel is minimized, but the bit error rate (BER) of the message channel is sufficiently low. Figure 5 depicts the impact of the message channel modulation on a 10 Gbit/s non-return-to-zero (NRZ)-modulated payload signal. At a BER of ${{1}}{{{0}}^{- 12}}$, a message channel modulation depth of 8.1% leads to a penalty of approximately 0.4 dB. As in G.698.4, the modulation depth $m$ is calculated from the average power values in the ones and zeros of the message channel, ${{P}}({{1}})$ and ${{P}}({{0}})$, respectively, as

 

Fig. 5. BER of a 10.3 Gbit/s payload signal over received power. A Manchester-encoded message channel at 50 kbit/s is envelope-modulated with different modulation depths onto the payload: solid line, no message modulation; dotted line, 5.3% modulation depth; dashed line, 8.1% modulation depth.

Download Full Size | PDF

 

Fig. 6. BER of a 50 kbit/s message channel in the presence of a 2.5 Gbit/s payload signal for different modulation depths between 3.6% and 5.8%.

Download Full Size | PDF

On the other hand, the BER of the message channel is reduced with increasing modulation depth. Figure 6 shows the BER as function of the received power for different values of the modulation depth. As the payload modulation acts as noise-like impairment for the message channel, a BER floor is observed, which is lowered with increased modulation depth. This BER floor is achieved at received power levels around ${-}{{40}}\;{\rm{dBm}}$, such that a reliable operation of the message channel is achieved already before the data channel operation is sufficiently established. It can be seen that for a modulation depth of more than 5.8%, the message channel error rate is below ${{1}}{{{0}}^{- 6}}$. A simple error correcting code, like a Hamming code, can be applied to yield a quasi-error-free message channel. In G.698.4, the modulation depth of the message channel was specified between 6.5% and 8%. A combination of (16,11) and (32,26) Hamming codes is applied on the 48-bit message frame.

C. Field Trials

In our previous research projects, we successfully demonstrated and evaluated the G.metro passive WDM fronthaul in a complete end-to-end 5G-ready mobile network. These demonstrations proved that the prototyped system not only fulfilled the stringent capacity and latency requirements of different use cases, but the solution also significantly helps service providers reduce the infrastructure cost and simplify the operation. A previous paper [24] summarized all the recent trials and outcomes of the proposed G.metro passive WDM system.

4. 100G OPTICAL INTERFACE FOR ETHERNET AGGREGATION

As mentioned in Section 2, another viable 5G fronthaul approach is to use high-speed Ethernet, such as 100 Gbit/s specified in the IEEE 802.3 working group, and to aggregate the signals of multiple RUs to leverage the statistic multiplexing gain. Low-cost and high-speed requirements lead to the considerations of using intensity modulation and direct detection (IMDD) together with gray optics for 100G trunk lines. Optical amplification should be avoided to again save cost and remove complexity, but in this way, the dispersion penalty also needs to be mitigated. Consequently, the O-band (1300 nm) transmission window becomes ideal to prevent severe limitations from chromatic dispersion. Among commercial solutions, four-lane 100G NRZ modulation was standardized as 100GBASE-LR4/ER4 in IEEE 802.3ba, where four parallel transceivers provide 25.78125 Gbit/s (64b66b line coded) each, operating in the O-band with a channel spacing of 800 GHz. The maximum reach without FEC is 10 km and 40 km for LR4 and ER4, respectively.

Although NRZ would be the simplest modulation format, this requires expensive high-bandwidth optics and electronics in multiple parallel lanes. Hence, we investigate a single-lane transceiver structure utilizing digital signal processing (DSP) for advanced modulation formats and FEC encoding on the host device in order to save the cost of optics. Among potential modulation formats, four-level pulse amplitude modulation (PAM-4) and discrete multi-tone (DMT) are under consideration because of their minimum DSP complexity.

Figure 7 shows the experimental setup used for evaluating two modulation schemes. Either real-time or offline DSP is applied at both the transmitter and receiver. It requires the use of high-resolution DAC and ADC components. At the transmitter, the differential outputs of the DAC are amplified by a linear driver with differential inputs and single-ended output, directly driving the electroabsorption-modulated laser (EML). The EML is temperature-stabilized, and the current of the distributed feedback laser (DFB) is properly set to ensure both the linear range and optical output power are around 1308 nm. The measured 3-dB bandwidth of this transmitter is around 27 GHz with a smooth roll-off. After transmission over up to 20 km standard single-mode fiber (SSMF) for PAM-4 and 22.4 km for DMT, the optical power is managed by a variable optical attenuator (VOA), and the signal is detected by a PIN-photodetector integrated with a linear transimpedance amplifier (PIN/TIA), resulting in a receiving bandwidth of 35 GHz. Finally, the signal is sent back to the ADC and stored for offline processing, or detected and processed in real time in the case of DMT.

 

Fig. 7. Experimental setup for 100G transmission with advanced modulation schemes.

Download Full Size | PDF

A. DMT Modulation and Signal Processing

DMT, as a special variant of orthogonal frequency division multiplexing (OFDM), employs the properties of Hermitian symmetry and the IFFT to create a real-valued signal with the frequency spectrum divided into orthogonal subcarriers. To achieve equal BER in all subcarriers in the presence of subcarrier-dependent signal quality, the modulation format in each subcarrier can be adapted, and its power can be allocated based on the water-filling method. This process is known as bit and power loading (BL, PL), and it enables the effective equalization of channel impairments and component bandwidth limitations without applying complex signal processing, e.g., a simple one-tap equalizer at the receiver side is sufficient for each subcarrier.

In the experiment, we use the Socionext real-time DMT DSP integrated circuit (IC) chip, which includes the DAC and ADC, and was mounted on an evaluation board to generate and analyze the DMT signal. The DMT chip can be configured to run at 64 GSample/s for a 100 Gbit/s data rate, which is of interest for this application, but other lower data rates are also possible. The signal consists of 255 subcarriers, where some are used as pilots for synchronization and equalization purposes. A cyclic prefix of 16 samples is used and the root mean square (RMS) related to the clipping ratio is dynamically optimized during the measurement. The net data rate corresponds to the Ethernet data rate, i.e., 103.125 Gbit/s for the 100G mode. The CI-BCH FEC [25] with a BER limit of 4.4e-3 is also implemented in the chip, giving post-FEC error-free results in the experiment. The chip has also implemented both BL and PL as well as adaptive background equalization to adapt slowly varying channels.

The receiving sensitivity is dependent on subcarrier modulation and data rate. Figure 8 shows the back-to-back BER performance under different data rates (25, 50, and 100 Gbit/s), where the 25 Gbit/s signal can be received error-free before FEC, while the 100 Gbit/s signal is still within the FEC limit, given a received power higher than ${-}{{6}}\;{\rm{dBm}}$.

 

Fig. 8. BER versus ROP for back-to-back transmission.

Download Full Size | PDF

The estimated signal-to-noise ratio (SNR) probed at a received power of ${-}{{3}}\;{\rm{dBm}}$ is shown in Figs. 9(a) and 9(b), characterizing the channel response under different data rates. Because of the zero-dispersion window in the O-band, the only impairment on the signal is the reduced signal power caused by fiber propagation loss, resulting in a lower SNR. This is also aligned with the BL results, as shown in Figs. 9(c) and 9(d), where only minor differences can be observed between the back-to-back and the fiber transmission cases. The maximum transmission reach is 22.4 km and 32.6 km at 100 Gbit/s and 50 Gbit/s, respectively.

 

Fig. 9. Measured results at different data rates in back-to-back and after maximum fiber transmission. (a)–(b) Estimated SNR versus electrical frequency, (c)–(d) bit loading, (e)–(f) power loading through amplitude scaling.

Download Full Size | PDF

In the O-band, no significant influence from the fiber transmission was observed, as long as the received power was above the optimum value of ${-}{{3}}\;{\rm{dBm}}$, as can be seen in Fig. 10. At 16 km distance, the maximum received power was ${-}{4.3}\;{\rm{dBm}}$, and the BER therefore started to degrade. It can be concluded that the system is only power-limited in the O-band, resulting in a reach of 22.4 km for 100 Gbit/s, which matches the expected upper limits of fronthaul reach.

 

Fig. 10. BER versus transmission distance at different data rates.

Download Full Size | PDF

All results were obtained in real time with a fully functional DMT IC chip, including FEC running stably in long-term measurements. Further details and similar performance using the same setup in the C-band can be found in [26].

B. PAM-4 Modulation and Signal Processing

PAM-4 encodes two bits into one symbol, resulting in a four-level signal and reducing the transmission bandwidth by a factor of 2 compared to on-off keying. Utilizing Nyquist pulse shaping with a small roll-off factor ($\beta \; \approx \;{0.1}$), the signal bandwidth can be further reduced, resulting in an electrical bandwidth of around 30 GHz for a 112 Gbit/s PAM-4 signal. Figure 11(a) shows the implemented offline DSP blocks for the Nyquist PAM-4 processing. A 4-ary deBruijn sequence of order eight (${{{4}}^8} = {{65}}{,}{{536}}\;{\rm{symbols}}$) was used and gray-mapped onto a PAM-4 signal. Compared to DMT, PAM-4 offers the possibility of easily compensating for the nonlinear transfer function of the modulator by adjusting the levels towards equally spaced power levels after the modulator. Afterwards, the signal was up-sampled to three samples/symbol, undergoing raised cosine shaping in the frequency domain with $\beta = {0.1}$, and down-sampled by a factor of 2, allowing the DAC to generate a 112 Gbit/s Nyquist-PAM-4 signal at 84 GSample/s. The 3-dB bandwidth of the DAC is 15 GHz.

 

Fig. 11. (a) DSP blocks of the PAM-4 system, (b) digital PSD of the transmit signal, (c) eye diagram after the EML.

Download Full Size | PDF

Furthermore, the subsequent digital pre-emphasis compensated for the bandwidth limitations of the DAC and driver, and the signal was quantized to use the full 8-bit resolution of the DAC. Figure 11(b) shows the power spectrum density (PSD) of the transmitted signal before the DAC with the effect of pre-emphasis (the gray area represents an uncompensated signal). Figure 11(c) illustrates the eye diagrams after the EML, exhibiting the typical over- and undershoots of a Nyquist PAM-4 signal. At the receiver, the signal was captured by the ADC, which ran at the same rate as the DAC and had a nominal bit resolution of 8 bits and a 3-dB bandwidth of 18 GHz. In the DSP, the signal was resampled to twofold oversampling. The clock recovery was based on the Gardner loop, and an adaptive symbol-spaced feed-forward equalization (FFE) is applied to recover the PAM-4 signal.

The performance of the pre-equalizer (Tx-FFE) in combination with the post-equalizer at the receiver (Rx-FFE) was evaluated with respect to the BER, in order to determine the necessary number of coefficients of both equalizers when the BER is below the FEC threshold. Figure 12(a) shows the difference of eye diagrams using either a Tx-FFE with 5 or with 61 coefficients. Figures 12(b) and 12(c) depict the BER against received optical power (ROP) for optical back-to-back transmission, using 5 and 61 Tx-FFE coefficients, respectively, in combination with different numbers of Rx-FFE coefficients. Applying 5 Tx-FFE coefficients and up to 31 Rx-FFE coefficients is necessary to achieve the BER below the CI-BCH FEC limit, while 11 Rx-FFE coefficients are required when 61 Tx-FFE coefficients are used. The relationship of numbers of coefficients between Tx-FFE and Rx-FFE is further illustrated as a contour plot in Fig. 12(d), where the achieved BER for different Tx-/Rx-FFE combinations is shown at a fixed ROP of 0 dBm. No significant BER improvement can be found with more than 11 Tx-FFE coefficients, while at the receiver at least 21 Rx-FFE coefficients are required. To achieve a BER below 1e-3, however, more than 40 coefficients for both Tx-FFE and Rx-FFE are necessary.

 

Fig. 12. Optical back-to-back transmission results of 112 Gbit/s PAM-4 employing different numbers of pre- and post-FFE coefficients. (a) Optical eye diagrams obtained directly after the EML using a pre-equalizer with 5 and 61 coefficients; (b), (c) BER versus ROP results using different numbers of post-FFE coefficients; (d) BER performance for different Tx-/Rx-FFE combinations at an input power of 0 dBm.

Download Full Size | PDF

Based on the results of Fig. 12, it can be concluded that the choice of 11 coefficients at Tx and 41 coefficients at Rx offers a good trade-off between performance and complexity. Applying this coefficient setting, the transmission performances of optical back-to-back, 10 km, and 20 km transmission are compared in Fig. 13. For back-to-back and 10 km SSMF, the results stay below the CI-BCH FEC threshold. However, the KP4-FEC threshold is not met. In addition, the limited output power of the EML prevents the possibility of transmitting over 20 km. Up to the achievable input power, a similar performance of the different transmission distances is shown. More details, including the result of partial-response PAM-4 can be found in [27].

 

Fig. 13. Transmission results of 112 Gbit/s Nyquist PAM-4 over different transmission distances using 11 Tx-FFE and 41 Rx-FFE coefficients.

Download Full Size | PDF

In summary, 100 Gbit/s transmission over 10 km of SSMF was successfully demonstrated for both real-time DMT and offline-processed PAM-4 if a CI-BCH FEC of 4.4e-3 is assumed, allowing error-free transmission. Using DMT even achieved a distance of 22.4 km. The overhead length of the FEC is 1 Mbit, leading to a processing latency of around 10 µs at 100 Gbit/s [25]. The other latency is primarily contributed from the number of equalizer taps for PAM-4 or the length of the (I)FFT for DMT. We found out that the other processing delay of the DSP is negligible compared to the one of FEC. Assuming a fiber propagation delay of 5 µs/km, the transmission of 10 km leads to an overall latency of around 60 µs.

Overall, the experimental results indicate a similar performance between DMT and PAM-4. However, the electrical and optical components could also dominate the actual performance.

5. TIME-SENSITIVE NETWORKING

Mobile networks towards 5G impose stringent requirements on high throughput, low PDV, and low latency due to emerging time-sensitive applications. As mentioned in Section 2, Ethernet is low-cost and flexible but lacking in determinism. Several projects for TSN are in progress [28], where the IEEE 802.1CM is a collaborative effort by the CPRI and IEEE organizations and defines Ethernet extensions for mobile fronthaul applications [29]. It specifies TSN mechanisms required to provide deterministic transport of 4G CPRI and 5G eCPRI streams facilitating reliable packet transport.

A. Timing in Ethernet

As originally designed, Ethernet treats all frames identically, regardless of the priority of traffic flows, which obstructs a direct adoption of Ethernet, especially for the fronthaul. As mentioned in Sections 1 and 2, this is because fronthaul is a LLS interface that is more sensitive to latency and PDV. Moreover, TDD-based wireless communication, multipoint radio access technologies, and cloud-RAN rely on precise time synchronization once adopted [30].

Although traffic classes using different priorities and virtual local area networks (VLANs) according to IEEE 802.1Q were introduced to separate traffic flows on the same LAN [31], mixing flows and port contention in Ethernet switches causes nondeterministic delay and thus degrades the quality of service of particular traffic. For instance, when the switch has already started the transmission of an Ethernet frame, while at that very moment, another frame (even with highest priority) also requests transmission on the same egress port, the latter frame has to be buffered until transmission of the previous one has been completed.

For the purpose of cell site synchronization, the most straightforward way would be to get the reference clock (typically, pulse per second, PPS) and time of day (ToD) from the global navigation satellite system (GNSS). Yet, this is costly for installation, and access to the GNSS satellite signal cannot be guaranteed at all times. As an alternative or addition, the frequency and time information can be distributed from one central source through the network. ITU-T specifies SyncE [32] to transfer a synchronization signal between network elements. It relies on a bit-level clock recovery solution similar to SDH/SONET and is highly robust for long-term frequency distribution, but it only provides frequency synchronization, and every node in the sync path requires physical-layer support for SyncE. Complementing SyncE, the IEEE 1588v2 PTP [11] provides a mechanism for end-to-end phase and time alignment, as well as frequency distribution if SyncE is not available. However, as a packed-based Layer 2/3 protocol, the performance of PTP can be extremely dependent on the PDV across network elements; indeed, ITU-T SG15/Q13 recommends use of SyncE and PTP for 5G fronthaul solutions in ITU-T Recommendation G.8275.1.

In the following sections, we review TSN solutions to 5G fronthaul and discuss an existing approach and ongoing work on TSN. We also present the proof of concept of a 100 GbE aggregator that prioritizes the time-sensitive service and ensure an accurate time synchronization for the TSN-enabled fronthaul.

B. TSN Solutions

The key to enabling determinism in TSN is the concept of sharing time. The IEEE 802.1 TSN task group selected a few of the most critical options from IEEE 1588-2008 to specify a profile, i.e., IEEE 802.1AS (and its revision 802.1AS-Rev), for time synchronization over a virtual bridged LAN (refer to IEEE 802.1Q). Since the standard Ethernet switching cannot avoid nondeterminism, many amendments or extensions to existing Ethernet standards were incorporated to enhance traffic scheduling and shaping [33].

IEEE 802.1Qbu [34] as part of the IEEE 802.1CM set of standards defines a pre-emption mechanism enabling minimized delay on deterministic (express) traffic when mixed with best-effort (pre-emptable) traffic within the same Ethernet ports. The transmission of low-priority (best-effort) packets is interrupted in order to yield to an intermittent higher-priority packet and resume the remaining transmission after the end of the high-priority packet. To do that, the two ends of the link need to activate the pre-emption support through the link layer discovery protocol (LLDP), and the Ethernet media access control (MAC) needs to be instructed to hold back the pre-emptable traffic. Frame pre-emption only works hop-by-hop and can be enabled only in networks that are aware of 802.1Qbu. 802.1Qbu specifies its bridge management, while 802.3br does the MAC part [35]. Since this scheme has a minimum fragment size of 124 octets in length, taking the mandatory interframe gap, preamble, and delimiter into account, an additional delay of 142 octets of transmission time might be incurred.

We evaluate the frame pre-emption technology and its behavior using a VHDL-based frame preemption model and running hardware behavioral simulations. We consider two types of traffic in the communication system: high-priority express traffic and low-priority best-effort traffic.

Figure 14 illustrates logarithmically the evaluation of the pre-emption delay behavior for express traffic by running a digital design hardware-based simulation using random packet lengths and packet gap sizes for both best-effort and express traffic. The simulation results confirm the calculated maximum PDV of 142 bytes of transmission time.

 

Fig. 14. Distribution of pre-emption delay for express traffic.

Download Full Size | PDF

C. FPGA-Based Implementation

A configurable FPGA implementation of a packet-based 100G Ethernet aggregator for 5G mobile fronthaul with low PDV using frame-pre-emption of best-effort traffic is demonstrated. We also present a P4 programmable FPGA-based IEEE 802.1Qbu compliant tester, which provides quantitative performance evaluation in terms of delay and PDV. The performance and viability of the proposed approach was shown in a live demonstration [36]. For demonstration purposes, a user interface running on a system controller monitors and configures both FPGA hardware platforms through the management interfaces. An overview diagram of our proposed testbed for demonstrating and evaluating a proof of concept of the fronthaul network is shown in Fig. 15. The fronthaul architecture enables eCPRI radio data transmission through a packet-based aggregation network, connecting RUs to the CU/DU.

 

Fig. 15. System overview of the TSN fronthaul architecture.

Download Full Size | PDF

 

Fig. 16. Detailed one-way data path architecture.

Download Full Size | PDF

The fronthaul signal from one CU/DU, together with standard PTP master clock messages, is emulated by the P4-based FPGA traffic analyzer and differentially forwarded either as express or best-effort data to the 10G port of the central office aggregator.

The system controller is connected to the management interface of each hardware platform, which facilitates the communication with the Microblaze System controller. The internal FPGA controller is implemented with support to on-board peripherals, such as PHY (Marvel 88E1111), DDR4, I2C, UART, and JTAG, as well as external peripherals SFP+ and QSFP28 plugs. This subsystem plays an orchestrator role. Its function is to configure and run systems, since the majority of processing is made by other subsystems implemented in VHDL and Verilog.

Figure 16 illustrates the one-way data path of the proposed implementation using two Virtex Ultrascale FPGA hardware platforms to evaluate packet-switched fronthaul traffic, addressing solutions for synchronization. One hardware platform is hosting the P4-based IEEE 802.1Qbu compliant traffic analyzer, which is optically connected to the second hardware platform hosting the two aggregator modules required for the proposed fronthaul implementation. The prototype platforms were designed with VHDL and C language for the Microblaze controller, and P4 language for packet processing functionalities on the Analyzer Platform, therefore being highly programmable. Manufacturer IP cores were also used.

The fronthaul signal from one CU/DU, together with standard PTP master clock messages, is emulated by the P4-based FPGA traffic analyzer and differentially forwarded either as express or best-effort data to the 10G port of the central office aggregator. The presented system is fully IEEE 802.1Qbu compliant on the client side 10 Gbit/s interfaces, while on the line side, the 100 Gbit/s interface we implemented is a strict priority (non-pre-emption) compliant arbitration mechanism. The MicroBlaze controller also maps the received messages at the hardware level into P4 specific match-action table abstraction for the generated P4 pipeline.

In the context of TSN, minimum latency and PDV measurements have been performed for express and best-effort data paths. This way, we measured a minimum latency of 1.16 µs on high-priority express traffic and 1.22 µs on low-priority best-effort traffic, while from the PDV point of view, we achieved a PDV of 224 ns on high-priority express traffic and 819 ns on low-priority best-effort traffic.

6. DEPLOYMENT OPTIONS

The preferred deployment solution for mobile backhaul will still be point-to-point or aggregated Ethernet. The option of using TDM-PON is technically feasible and complementary to FTTx deployments. Yet, PON solutions are not widely used for practical reasons [37]. Nonetheless, XG(S)-PON can still be considered as the next deployable solution for the mobile backhaul [38]. Such a solution offers 10 Gbit/s downstream and 2.5 or 10 Gbit/s upstream on a paired wavelength channel, and it is able to coexist on the same optical distribution network, saving the cost of the fiber infrastructure. The evolution to the next-generation high-speed PON (i.e., 25G/50G on a single wavelength) is still under discussion within ITU-T and IEEE.

In contrast to the backhaul, the fronthaul (as well as the midhaul) introduced by the 5G RAN makes the optical fiber transport more critical, as the LLS requires more capacity and stricter latency margin. In light of eCPRI adoption, for instance, the data rate per interface port is ramping up from 10 to 25 Gbit/s, and the maximum one-way frame delay under the highest class-of-service is only 25 µs for ultralow latency applications. Although point-to-point dark fiber using gray optics is the most straightforward and cheapest solution in terms of equipment, fiber availability in some regions and scaling to massive multiple input–multiple output (MIMO) are questionable.

Therefore, given Ethernet encapsulated fronthaul traffic, high-speed Ethernet interfaces (e.g., 100 GbE or higher) can offer more capacity with statistical multiplexing gain. Advanced optical modulation techniques (e.g., DMT and PAM-4) proposed in Section 4 proved that the 100 Gbit/s interface can be achieved with a single gray transceiver at low cost. Also, low-cost coherent 100 Gb/s transceiver technology could play a role in the future here. Having said that, aggregation along with FEC and packet processing introduces a latency of tens of microseconds, and scaling to higher line rate (e.g., 400 GbE) is still under development and immature. To further exploit the fiber resource while keeping the simplicity of the dark fiber, the point-to-point passive WDM on a single fiber working will be the ultimate solution in areas where fiber is scarce, since dedicated wavelengths for each CU/DU-RU pair are equivalent to a dedicated fiber link (“virtual fiber”). Such a system is also easy to upgrade as it grows. Although traditionally WDM is considered an expensive approach due to either complex inventory of fixed optics or costly tunable optics, the G.metro wavelength-agnostic WDM solution employing the low-cost tunable transceivers can lower the cost barrier and enable remote management capability.

As for the data transport protocol, Ethernet becomes the convergence mechanism and needs to be extended to support the TSN in order to meet fronthaul latency, timing, and synchronization requirements. While OTN [39] and FlexE [40] solutions (with tributary granularities of 1.25 and 5 Gbit/s, respectively) could also be used for fronthaul, both of them lack statistical multiplexing gain and introduce an additional networking layer that can become an impediment to future network upgrades.

 

Fig. 17. Layer stack for 5G fixed transport.

Download Full Size | PDF

Putting all these facts together, Fig. 17 summarizes all the layering options for 5G-RAN transport. It is worth noting that there is no one-fits-all transport solution. Depending on the fiber infrastructure and network requirements, different Ethernet-based transport and optical underlay solutions can be applied, either standalone or in combination.

7. CONCLUSION

The 5G-RAN architecture evolved introducing different FSs. Rather than concentrating all the network intelligence in a centralized location, as in 4G/LTE C-RAN with CPRI, 5G allows a more flexible allocation of radio protocol stack functions between the CU, DU, and RU. The LLS fronthaul interface splits are most demanding in capacity, latency, and timing requirements.

Ethernet technology is considered in the 5G standards to consolidate both front- and backhaul traffic on a common optical network infrastructure. G.metro passive WDM access using the low-cost tunable transceiver offers scalable wavelength as service between the CU/DU and RU to improve fiber link efficiency while retaining lower cost compared to a traditional WDM solution. We also demonstrated that advanced modulation formats like DMT and PAM-4 can enable 100 Gbit/s transmission between Ethernet aggregation sites using low-cost gray optics.

On top of these transport media, TSN adds the capability of controlling the packet delay and delay variation of Ethernet, which is necessary for the radio protocol. Our TSN implementation based on IEEE 802.1CM was demonstrated to be promising for delivering time-sensitive fronthaul traffic.

Overall, 5G-RAN transport requires improvements in both capacity and latency control and needs to be optimized depending on the actual deployment.