1. Introduction

As traffic in intra-datacenter networks (DCN) continues to grow [1], there is strong interest in increasing per-fiber data channel counts in order to scale data rates while reducing cabling footprint [1–3]. In addition to coarse and dense wavelength division multiplexing (CWDM/DWDM) in the 1.3μm and 1.5μm bands, there is increasing interest in space division multiplexing (SDM) for intra-DCN [2–4]. With SDM, independent data channels can be transported over multiple spatial cores, modes or elements of a single fiber. However, for few-mode fiber SDM, coherent detection and computationally intensive multiple input multiple output (MIMO) digital signal processing (DSP) have previously been required due to random mode mixing during transmission [5–7]. Eliminating both coherent detection and MIMO DSP would enable the use of commercial direct detection transceivers, such as small form-factor pluggable SFP + transceivers, making few-mode fiber SDM attractive for intra-DCN.

In this paper, to the best of our knowledge, we experimentally demonstrate the first few-mode SDM transmission of real-time 10Gb/s Ethernet (10GbE) traffic using commercial SFP + transceivers with no coherent detection or MIMO DSP using all the spatial modes of a 0.5km elliptical-core few-mode fiber (EC-FMF). Three-mode (LP₀₁, LP_11e, LP_11o) × 10GbE transmission in the 1.3μm band and dual-mode (LP₀₁, LP_11e) × 10λ × 10GbE/λ transmission in the 1.5μm band are experimentally evaluated in terms of both real-time throughput and bit error rates (BER). We achieve modal crosstalk <-26dB for the supported modes in the 1.3μm and 1.5μm bands. The very low crosstalk between the normally degenerate LP_11e and LP_11o modes is enabled by the EC-FMF design and is a novelty compared with previous work [5–8]. Low crosstalk between LP₀₁ and LP_11e modes is moreover verified over 2km EC-FMF at 1.5μm using 100Gb/s coherent dual polarization (DP)-QPSK transmission. By enabling spatial channel density gains, the approach is attractive for future intra-DCN communication.

2. SDM intra-DCN architecture featuring elliptical-core few-mode fiber

Figure 1(a) shows the proposed SDM intra-DCN architecture based on the architecture in [2] but featuring an EC-FMF [Fig. 1(b)]. The DCN contains m clusters of n interconnected racks, with each rack comprising tens of servers. Servers are interconnected with top-of-racks (ToRs) and ToR 10GbE optical transceivers (Tx/Rx) use fixed-λ lasers in the 1.3μm or 1.5μm bands. Spatial multiplexers/demultiplexers (S-MUX/S-DEMUX) combine signals from ToR transceivers for transmission over the EC-FMF to/from different ToRs and/or clusters. Inter-cluster transmission may require higher data rates (e.g. 100Gb/s), as shown in Fig. 1. Connectivity between ToRs and/or clusters is achieved by the reconfigurable optical cross-connections (OXC) [1, 2]. We note that optically-connected electrical switches currently dominate, while reconfigurable OXC are limited to specialty scenarios and might play a larger role in the future. In conventional SDM using few-mode fibers with non-elliptical (circular) cores, propagation constants along fiber core axes are approximately equal for all spatial modes in a mode group due to fiber core symmetry. Consequently, random mode coupling (crosstalk) between modes arises, necessitating the use of MIMO DSP to undo mode coupling and modal dispersion [5–7]. However, in EC-FMF [Fig. 1(b)], the different mode groups have different propagation constants due to birefringence arising from core asymmetry (ellipticity). As a result, the normally degenerate LP₁₁ modes do not couple even under fiber twisting/bending, such that end-to-end modal crosstalk is dominated by the S-MUX/DEMUX. The spatial channels can thus be considered non-interfering, requiring only S-MUX/DEMUX with low crosstalk. Compact, low-crosstalk S-MUX/S-DEMUX can be made via core-to-core coupling [9] or by tapering SMF connectors into the EC-FMF using a mode-selective photonic lantern [10]. Unlike specialty fibers for SDM, the EC-FMF of Fig. 1 is made using a conventional fiber manufacturing process. These advantages, in addition to cabling footprint reduction and compatibility with commercial transceivers render the proposed SDM approach highly attractive for intra-DCN spatial channel density gains. To design a fiber that supports non-degenerate LP modes, it is necessary to break the circular symmetry. One way to do so is to make the fiber core elliptical since this breaks the degeneracy of modes in a mode group of a circular core fiber. Figure 1(c) shows the EC-FMF design. A low index ring can be added next to the elliptical core to improve performance under bending. We define an ovality parameter, χ, to describe the elliptical core shape as

 

Fig. 1 (a) SDM intra-DCN architecture; (b) elliptical-core few-mode fiber (EC-FMF); (c) EC-FMF design; (d) Effective indices for three mode groups.

Download Full Size | PDF

Figure 2(a) plots the average effective index difference $\Delta n_{eff}{,}_{11}$between the two LP₁₁ modes. The effective index difference increases with the core ovality. For χ > 30%, $\Delta n_{eff}{,}_{11}$> 10⁻³ between the two LP₁₁ modes, which is large enough to reduce the mode coupling between them. We note that $\Delta n_{eff}{,}_{11}$≤ 10⁻⁶ between the two polarizations of the same LP₁₁ mode. Figure 2(b) shows the EC-FMF cross-section. The two dark regions are stress-applying parts for creating asymmetric stress during fiber draw process to make the core elliptical. The R_a ~15 μm, and R_b ~10μm. A core aspect ratio ~1.5 is achieved, which corresponds to χ = 40%. The elliptical core design features a graded index profile with α ≈2 along both axes to minimize the differential group delays among the modes. The EC-FMF was made using outside vapor deposition and a conventional fiber draw process.

 

Fig. 2 (a) Effective index difference between the two LP₁₁ modes; (b) EC-FMF cross-section.

Download Full Size | PDF

3. Experimental setup and results

Figure 3 shows the experimental setup for EC-FMF transmission in the 1.3μm [Fig. 3(a)] and 1.5μm [Fig. 3(c)] bands, respectively. In both cases, real-time 10GbE data traffic with 64b/66b line coding was generated in a Dell PowerEdge R420 server (Server 1) and sent via a 10GbE switch over 0.5km EC-FMF to a second Dell server (Server 2) configured to measure real-time traffic throughput. In both bands, commercial SFP + transceivers with no forward error correction (FEC) or polarization mode dispersion (PMD) compensation were used for real-time 10GbE traffic transmission. The return SMF loopback link from Server 2 to Server 1 was used for TCP-layer link establishment and acknowledgments. At 1.3μm, the EC-FMF supports LP₀₁, LP_11,e and LP_11,o modes, while at 1.5μm, LP₀₁, and LP_11,e modes are supported. The propagation loss of the EC-FMF was measured to be about 0.46dB/km at 1.3μm for all three supported modes, and 0.25 dB/km at 1.5μm for both supported modes. For the 1.3μm experiment [Fig. 3(a)], three independent real-time 10GbE data streams from three SFP + transceivers were spatially multiplexed and transmitted over the EC-FMF. The S-MUX/DEMUX setup is shown in Fig. 3(b), and uses phase plates for spatial modulation. For the 1.5μm experiment [Fig. 3(c)], a 10λ × 10Gb/s/λ WDM signal was transmitted over each spatial mode. Ten DFB lasers at frequencies f_n = (192.10 + 0.2n) THz, n = 0, 1, .., 9, were multiplexed and intensity modulated (IM) with 10Gb/s on off keying (OOK) signals. Data on WDM channels were de-correlated via two interleavers. A wavelength selective switch (WSS) was used to remove one of the DFB channels and replaced with the real-time 10GbE data channel from the SFP + transceiver of the same f_n under test. The WDM signal was split into two copies; one copy was transmitted over LP₀₁ and the other transmitted over LP_11e after a τ_d = 92.1ns de-correlation delay. Measurements were repeated for ten 1.5μm SFP + transceivers with center frequencies f_n. Figure 3(d) shows the 3 × 3 coupling matrix measured at 1.3μm while Fig. 3(e) shows the 2 × 2 coupling matrix at 1.5 μm. The diagonal elements are the insertion losses for the three modes, while the off-diagonals represent crosstalk. Insertion loss is dominated by passive beam splitters [Fig. 3(b)], and can be greatly reduced via S-MUX/DEMUX methods in [9, 10]. Figures 3(d)-(e) show that the crosstalk at 1.3μm is between −26.9dB to −29.3 dB, and crosstalk at 1.55 μm is between −28.5dB to −29.4dB. The total loss was measured at 1.3μm under several fiber bending scenarios. The differential mode group delay (DMGD) at 1.3μm was measured to be 1.28ns/km between LP₀₁ and LP_11e (LP_11e arrives earlier) and 2.04ns/km between LP₀₁ and LP_11o (LP_11o arrives earlier). At both 1.3μm and 1.5μm, system performance was evaluated in terms of real-time 10GbE traffic throughput and BER vs. received power, which was measured using data captures by a real-time oscilloscope placed at the S-DEMUX output, followed by offline processing of 25 million bits per BER point. Finally, 100Gb/s DP-QPSK transmission at 1.5μm with coherent reception and MIMO DSP was evaluated over 2km EC-FMF. In this case, the 100Gb/s Tx and Rx were placed at the S-MUX input and S-DEMUX output, respectively.

 

Fig. 3 Experimental setup: (a) 1.3μm experiment; (b) spatial multiplexer and demultiplexer (S-MUX/S-DEMUX); (c) 1.5μm experiment; (d) 3 × 3 coupling matrix at 1.3μm; (e) 2 × 2 coupling matrix at 1.5μm.

Download Full Size | PDF

Figure 4 shows results for the 1.3μm experiment. Real-time 10GbE throughput was sustained at ~9.9Gb/s for all three modes [Fig. 3(a)] over the observation interval, confirming system stability and an operational BER <10⁻¹² without FEC. A ≤ 2dB penalty at BER = 10⁻⁶ was achieved compared to optical back-to-back (BTB), as per Fig. 4(b).

 

Fig. 4 Experimental results, 1.3μm: (a) real-time throughput; (b) BER vs. Rx power, 10GbE; (c) total loss (dB) under fiber bending scenarios and LP₀₁ launch; (d) total loss (dB) under fiber bending scenarios and LP_11e launch; (e) total loss (dB) under fiber bending scenarios and LP_11o launch.

Download Full Size | PDF

Figures 4(c)-(e) show the total loss under several fiber bending scenarios for LP₀₁, LP_11e, and LP_11o launch, respectively. As shown in Figs. 4(c)-(e), bend losses for the LP_01,e are ~4, 5, 7 and 12 dB/loop for bend radii of 23, 18, 16 and 14 mm, respectively, while the bend losses for the LP_11,o are slightly lower at ~1, 3, 5 and 10 dB/loop for the same bend radii set. Figure 5 shows results at 1.5μm, confirming stable ~9.9Gb/s real-time 10GbE throughput for both modes [Figs. 5(a)-(b)], and ≤ 1.5dB penalty at BER = 10⁻⁶ due to PMD compared to optical BTB for all ten λ [Fig. 5(c)]. As shown in Fig. 5(d), 2 × 2 MIMO (polarization de-multiplexing only) and 4 × 4 MIMO (polarization + spatial de-multiplexing) were virtually identical for both modes, showing that the LP₀₁ and LP_11e can be regarded as non-interfering channels by commercial 100Gb/s coherent DP-QPSK transceivers.

 

Fig. 5 Experimental results,1.5μm: (a) Iperf screenshot of real-time throughput; (b) real-time throughput; (c) BER vs. Rx power, 10GbE; (d) BER vs. OSNR, 100Gb/s coherent QPSK.

Download Full Size | PDF

4. Summary and conclusions

We have demonstrated the first few-mode SDM transmission of real-time 10GbE traffic using commercial SFP + transceivers without coherent detection or MIMO-DSP over 0.5km EC-FMF, achieving <-26dB crosstalk between LP_11e and LP_11o modes at 1.3μm. Low crosstalk was moreover verified between LP₀₁ and LP_11e modes over 2km EC-FMF at 1.5μm via 100Gb/s DP-QPSK transmission. By supporting channel density increases via low-crosstalk SDM, the approach is highly promising for future intra-DCN communication.