Data center elastic wireless optical interconnection system based on cascade connection of RMS and TMS

Abstract

The invention relates to wireless optical interconnection of a metasurface and a data center, in particular to a data center elastic wireless optical interconnection system based on cascade connection of an RMS and a TMS, which comprises a reflection metasurface and a transmission metasurface. A wave beam output by a first optical fiber array port arranged at the top of the source cabinet outputs N wave beams with different azimuth angles and emergence angles xi through the transmission metasurface, the N wave beams are transmitted to the corresponding second optical fiber arrays respectively, and the wave beam output by each second optical fiber array is modulated by the LCPC and deflected and focused by the lens. 2N different wave beams are generated through the reflection metasurface to cover the cabinet; according to the method, at least 2 * N times of wave beams of a traditional SLM or MEMS cover the area of the equipment cabinet, the probability of successfully establishing a communication link between the equipment cabinets is greatly increased, in addition, rotation of the RMS can be controlled according to the flow prediction result, and the topological scale of the data center elastically changes according to the time-varying flow.

Description

Technical Field

[0001] The present invention relates to wireless optical interconnection between metasurfaces and data centers, and particularly to an elastic wireless optical interconnection system for data centers based on a Reflective Metasurface (RMS) and a Transmissive Metasurface (TMS). Background Art

[0002] Data Centers (DCs) carry bandwidth-intensive parallel computing services such as large language models, which pose challenges to the electrical switching rate supporting inter-rack communication. In particular, the number of ports and design complexity of electrical switching units will show explosive growth. Existing wireless optical switching units include Spatial Light Modulators (SLMs) and Micro-Electro-Mechanical Systems (MEMS), which can break through this bandwidth bottleneck. They "guide" the beam from the server port of a certain rack to the server ports of other racks, establishing a laser link between racks with low inter-channel crosstalk. However, there are some technical problems with current mainstream wireless optical switching units: controlling the beam steering of the SLM by adjusting the electronic phase can only provide a limited steering angle (~3°); MEMS can achieve more flexible beam steering by injecting an external voltage to adjust the rotation of the micromirror, but a large number of steering elements need to be configured for multi-beam operation, and the control is complex. Moreover, both SLM and MEMS only support one-to-one communication, and high-bandwidth switching requirements can only be met at the cost of cumbersome system control and configuration. Summary of the Invention

[0003] In order to enable communication between a source rack and more racks, significantly expand the DC topology scale, and improve the DC switching capacity, the present invention proposes an elastic wireless optical interconnection system for data centers based on the cascading of RMS and TMS, including a reflective metasurface and a transmissive metasurface, to establish a communication link between a source rack and multiple racks. The beam output from the first fiber array port set on the top of the source rack passes through the transmissive metasurface and outputs N beams with different azimuth angles and the same exit angle of ξ, which are respectively transmitted into the corresponding second fiber arrays. The beam output from each second fiber array is modulated by LCPC, deflected and focused by a lens, and then 2N different beams are generated through the reflective metasurface to cover the racks.

[0004] Further, N depends on the chip size of the transmissive metasurface and the lattice constant of the on-chip unit structure.

[0005] Further, the maximum value of N is 2×l m , l m$N$ is the maximum number of OAM channels supported by the chip of the transmissive metasurface, and its calculation method is as follows:

[0006]

[0007] Among them, $W$ represents the chip size of the transmissive metasurface, and $\sigma$ represents the lattice constant of the on-chip unit structure of the chip of the transmissive metasurface.

[0008] Furthermore, the chip size of the transmissive metasurface is 45μm×45μm, the lattice constant $\sigma$ of the on-chip unit structure of the chip of the transmissive metasurface is 950nm, and the maximum value of $N$ is 6.

[0009] Furthermore, the chip of the transmissive metasurface is an array composed of multiple units, and each unit consists of a SiO 2 substrate and Si nanowires, and the Si nanowires are located at the center of the SiO 2 substrate.

[0010] Furthermore, the thickness of the Si nanowires is 1.2μm.

[0011] Furthermore, the system also includes a traffic prediction model, which controls the rotation of the reflective metasurface according to the prediction result of the prediction model, so that the topology scale of the data center changes elastically with the time-varying traffic, avoiding the situation of service denial caused by excessive traffic or resource allocation redundancy caused by too little traffic under the static topology.

[0012] Compared with the prior art, the present invention expands the coverage area of the beam to the cabinet, enables the source cabinet to establish communications with more cabinets within the coverage area, significantly expands the DC topology scale, improves the DC switching capacity, and at the same time can control the rotation of the RMS according to the traffic prediction result, so that the DC topology scale changes elastically with the time-varying traffic. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 Schematic diagrams of the RMS and TMS of the multi-beam modulation ability of the present invention;

[0014] Figure 2 Schematic diagram of the general radio-optical data center architecture based on TR-MS of the present invention;

[0015] Figure 3 Side view and top view of the nano-unit structure of the TMS chip of the present invention;

[0016] Figure 4 Schematic diagram of the angle rotation in the present invention;

[0017] Figure 5 Light intensity distribution diagram and corresponding optical field distribution diagram of the TMS chip of the example method of the present invention under the normal incidence of linearly polarized light;

[0018] Figure 6 Comparison chart of the prediction performance of the En-LSTM and the classical LSTM of the present invention;

[0019] Figure 7 Process diagram of the AI-assisted DC topology reconstruction based on the present invention. Detailed implementation manners

[0020] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0021] The present invention proposes a data center elastic wireless optical interconnection system based on the cascade of RMS and TMS, including a reflective metasurface and a transmissive metasurface, establishing a communication link between a source cabinet and multiple cabinets. The beam output from the first fiber array port provided at the top of the source cabinet passes through the transmissive metasurface and outputs N beams with different azimuth angles and the same exit angle of ξ, which are respectively transmitted to the corresponding second fiber arrays. After the beams output from each second fiber array are modulated by LCPC and deflected and focused by a lens, 2N different beams are generated through the reflective metasurface to cover the cabinets.

[0022] In the present invention, the TMS generates N output signals after acting on the incident light beam. Each output signal is transmitted through LCPC, Half lens, and RMS. Each RMS is used to generate a reflective coverage area with a size of L x ×L y The cabinets are responsible for receiving the light beams to establish a communication link. In the present invention, both the TMS and the RMS operate at a wavelength of 1550 nm. Among them, the TMS acts as a splitter to increase the number of its channels N as much as possible before the light beam passes through the RMS. The RMS is used as a reflective device to achieve multiple relatively wide reflective coverage areas, and the number of specific coverage areas is equal to N.

[0023] The existing RMS has good parallel beam control capabilities. Such as Figure 1(a), the fiber optic array located at the top of the source cabinet has a scale of m×n. The optical signals emitted by these light beams are first modulated by a liquid crystal polarization controller (LCPC), and then deflected by a half lens (Half lens) and reflected by the RMS respectively. Since a plane wave has two polarization states: x polarization and y polarization. For the incidence of the x-polarized beam, the RMS acts as an ordinary mirror, and the reflection angle is equal to the incidence angle. For the incidence of the y-polarized beam, due to the phase gradient on the surface of the RMS, an extraordinary reflection beam is generated, and the extraordinary reflection angle and the incidence angle satisfy the generalized Snell's law. In the figure, the area S of the coverage area of the extraordinary reflection A is much larger than that of the normal reflection S N , then S N +S A >2×S N , at least twice the reflection beam coverage area of the SLM or MEMS. In principle, the source cabinet can establish communication with any other cabinet within the coverage area.

[0024] As Figure 1 (b), the TMS has the ability of one-to-N communication. After the linearly polarized light is transmitted through the TMS, N beams are output along different directions. Among them, N depends on the chip size of the TMS and the lattice constant of the on-chip unit structure. After the N output beams are reflected by their respective RMSs, theoretically, it can form an area of at least 2×N times the beam coverage of the SLM or MEMS to realize communication between the source cabinet and more cabinets, thereby significantly expanding the DC topology scale and improving the DC switching capacity. Figure 3 This is the basic nano-unit structure diagram of the TMS chip designed by the present invention, which consists of a SiO 2 substrate and Si nano fins, where the height H of the nano fins is 1.2 μm. Through the combined action of the off-axis design and the spin Hall effect, the TMS chip can independently separate and control the OAM and SAM of different beams, thereby realizing one-to-N communication. Theoretically, a beam carrying different OAMs has infinitely many mutually orthogonal phase wavefronts, which can be characterized by the topological charge l (l = 0, 1, 2,...); a beam carrying different SAMs has mutually orthogonal circular polarization states, including left-handed circular polarization (LCP, s = 1) and right-handed circular polarization (RCP, s = -1). Through the combined action of the off-axis design and the spin Hall effect, the OAM and SAM of different beams can be independently separated and controlled, thereby realizing one-to-N communication. The maximum number of OAM channels l supported by the TMS chip designed by the off-axis method is calculated as follows: m Calculated as follows:

[0025]

[0026] Among them, W represents the chip size of the transmissive metasurface, and σ represents the lattice constant of the unit cell structure on the chip of the transmissive metasurface. Since there are two circular polarization states of left and right handedness, the maximum value of N is 2·l m The designed TMS chip size of the present invention is 45μm × 45μm, and the lattice constant σ = 950nm is set. According to the above formula, the l of this TMS chip can be estimated m = 3, which means that a communication mechanism with N = 6 can be supported. For simplicity and without loss of generality, the present invention only considers 4 channels (N = 4, t = 2) with different angular momentum states (|s, l|, where s = ±1, l = 1, 2).

[0027] Essentially, to achieve one-to-N communication, the key lies in designing the phase profile of the TMS surface. The interaction process between the incident light and the TMS can be characterized as:

[0028]

[0029] Among them, LCP represents left-handed circular polarization; RCP represents right-handed circular polarization; is the action symbol; i represents the imaginary unit; l tL represents the topological charge carried by the left-handed circularly polarized light; l tR represents the topological charge carried by the right-handed circularly polarized light; θ represents the azimuthal coordinate of the transmitted beam; ρ represents the transverse position vector of the transmitted beam; f represents the focusing position of the transmitted beam, k 0 is the transverse wave vector in vacuum; k tL and k tR are the transverse wave vectors of the transmission function; t is the output channel number index corresponding to each circular polarization state, and its value range is [0, l m ; x represents the x-axis coordinate of the Si nanowire fin; y represents the y-axis coordinate of the Si nanowire fin;.

[0030] In this embodiment, the total modulation phase of the TMS surface is defined as:

[0031]

[0032] Among them, φ L represents the left-handed circular polarization modulation phase; φ R represents the right-handed circular polarization modulation phase; ang{·} is a function for calculating the complex angle.

[0033] The TMS chip designed in the present invention can not only endow the incident light with an independent phase, but also make the transmitted beam output carry opposite circular polarization states and the expected topological charge. This process can be achieved by the combination of geometric (PB) phase and dynamic phase. The former is related to the rotation angle of the on-chip nano-fin, and the latter mainly depends on the size of the nano-cell and the transmittance of the nano-fin material. In addition, the incident light in this paper is linearly polarized light, which can be regarded as the superposition of LCP and RCP, then the PB phase and the dynamic phase are calculated as follows:

[0034]

[0035] To effectively control the and of the incident light, the unit structure parameters of TMS should be strictly designed.

[0036] The interconnection scheme of the present invention operates in the general wireless optical data center architecture of TR-MS. As Figure 2 shown, this architecture is introduced with N = 2 as an example in this embodiment. An m×n fiber optic array (purple) is placed on top of the source cabinet, and each array port corresponds to a TMS. The beam emitted from one fiber optic array port generates N = 2 beams with different azimuths and the same emission angle ξ after passing through the TMS, and is coupled into an m×n fiber optic array (red) at an angle ξ with these beams in the x direction, ensuring that it is then accurately projected onto their respective RMSs; subsequently, the beams output from the fiber optic array are each modulated by LCPC, their polarization states are determined, and then deflected by a half mirror and reflected by the RMS. Since there is a phase gradient on the surface of the RMS in the x direction, the x- and y-polarized incident beams will exhibit different phase responses, and normal and abnormal reflections will occur simultaneously, finally forming 2*N = 4 reflected beams to cover the cabinet, thus determining the initial topological scale of the DC in this example. In the solution of the present invention, the source cabinet can establish communication with any cabinet within the beam coverage area. If the initial topological scale is fixed, the number of other cabinets with which the source cabinet successfully establishes communication within the coverage area will also be fixed. Then, a lower traffic volume in a certain time slot will result in a larger resource redundancy overhead. On the contrary, due to the too small topological scale, the source cabinet cannot establish communication with cabinets farther away and is denied service. In fact, the DC elephant flow changes according to certain rules and has characteristics such as self-similarity and periodicity. Based on this consideration, the control unit predicts the bandwidth demand of the elephant flow in a future time slot. If there is an imbalance with the highest network bandwidth supply under the current topological scale, it calculates a topological scale that just serves the predicted traffic bandwidth demand, and thus quickly rotates the RMS at the desired angle to achieve DC topological reconstruction with the change of traffic.

[0037] The present invention also designs an AI-assisted traffic prediction model that fully considers the self-similarity and periodicity of elephant flows. Based on this model, the DC elastic topology reconstruction algorithm can control the rotation of the RMS according to the traffic prediction results, enabling the DC topology scale to change elastically with time-varying traffic, and avoiding situations such as service denial due to excessive traffic or resource allocation redundancy due to insufficient traffic under a static topology. As an alternative implementation, the data center ETR algorithm based on the prediction model can control the two-dimensional rotation of the RMS according to the aforementioned traffic prediction results, expanding / cropping the area of the coverage region, enabling the data center topology scale to change elastically with time-varying traffic, and avoiding situations such as service denial due to excessive traffic or resource allocation redundancy due to insufficient traffic under a static topology.

[0038] In Figure 3 (a), the on-chip unit structure of the chip of the transmissive metasurface is given, which consists of an SiO 2 substrate and Si nanowires (Palik model), where the height H of the nanowires is 1.2 μm, and the working wavelength of the TMS is 1550 nm, which is the same as that of the RMS chip. Here, the value of H is set to be close to the working wavelength, and the TMS communication performance is the best in the case of 1.2 μm in the simulation scan. Since the geometric structure of the nanowires is asymmetric, it will exhibit birefringence similar to that of a rectangular waveguide. As Figure 3 (b), by adjusting the length X and width Y of the nanowires, the corresponding phase shift (along the x direction) and (along the y direction) can be controlled. In this embodiment, the scanning ranges of X and Y are from 150 nm to 780 nm, the lattice constant σ is set to 950 nm, periodic boundary conditions (PBC) are adopted in both the x and y directions, and a perfectly matched layer (PML) is adopted in the propagation direction z. In this embodiment, the FDTD method is used to calculate the transmittance and phase shift of nanowires with different geometric parameters. Figure 3 (c) and (d) respectively give the transmittance changes of nanowires with different X and Y values when linearly polarized light is incident along the x and y directions, Figure 3 (e) and (f) respectively give the corresponding phase shift changes. The results show that and can independently cover the phase range of 0 to 2π, and at the same time, the transmittance can reach more than 80%. Therefore, by correctly selecting the size and rotation angle α of the nanowires, efficient manipulation of and can be achieved. As an alternative implementation, those skilled in the art can select 8 nanowires that meet the requirements as an eight-order phase unit of the TMS according to actual needs. Additionally, combining Figure 3 (g), the and The phase difference between adjacent nanoscale fins is close to π. At the same time, the phase step between adjacent nanoscale fins is about π / 4, which is basically consistent with the theoretical step of the eighth-order phase unit. Therefore, matches the eighth-order theoretical phase value. As an alternative implementation, when , then at this period, 0.5 / 0.125 (average step), that is, the 4th nanoscale fin, is selected, and the rotation angle of the nanoscale fins required at different periods can be determined. And it should be noted that the size parameters of the 8 nanoscale fins are all different.

[0039] In this embodiment, the power conversion efficiency η of the nanoscale fins is calculated as η = P o / P i > 90%, that is, the insertion loss is as low as 0.5 dB, where P i is the incident beam power, and P o is the transmitted beam power. At the same time, a higher power conversion efficiency means that the incident light can still maintain a high transmission quality at the output end after passing through the TMS, and it is fully capable of cascading with the RMS (insertion loss is 3 dB). To simplify the calculation, this embodiment ignores the single-mode fiber loss during the transmission of the beam from the TMS to the RMS. Then the total insertion loss is about 3.5 dB. The size of the TMS chip designed in this embodiment is 45 μm × 45 μm. According to the calculation formula of l m , l m can be calculated as l

[0040] = 3, which means that it can support 1 pair of N = 6 communications.

[0041] In addition, to more reasonably manage traffic allocation, this embodiment proposes the traffic demand in the next time slot predicted by the En-LSTM model. If it significantly exceeds the highest network bandwidth supply under the current DC topology scale, then the RMS should be rotated without disturbing the previous communication between cabinets to obtain a wider beam coverage area, so that the source cabinet can successfully establish communication with more new cabinets and improve the highest network bandwidth supply of the DC.

[0042]

[0043] Among them, a is an adjustable experimental parameter. In this embodiment, a = 0.7 is obtained after multiple delay screenings; b is a random number uniformly distributed from 0 to 1; X t b represents the best goshawk individual at the t-th iteration, and X t+1i represents the i-th goshawk individual at the (t + 1)-th iteration. The total number of goshawk individuals is N, and T is the total number of iterations. For each individual, the original search strategy or the shrinking spiral search strategy is randomly selected in each update. In this implementation, the goshawk individual is the number of neurons, the learning rate, and the number of training times in the LSTM model used in the present invention for predicting traffic.

[0044] As Figure 4 , the point P corresponds to the normal reflection landing point of the left-edge light at x 1 , and the point E corresponds to the abnormal reflection landing point of the right-edge light at x m . When the RMS rotates counterclockwise around the axis, both points P and E will move to the right, and vice versa. When the point P moves out of the right boundary of the Rack 1 (i.e., |OP| > g, where g represents the length of a single cabinet), the original communication between the Rack 1 and other cabinets will be interrupted. Therefore, on the premise of considering not interrupting the previous communication between cabinets, the exhaustive method is used to obtain the set Φ of topological patterns formed during the rotation of the RMS, and the set β of rotation angles of the relevant RMS under each topological pattern. Since the rotation of the RMS only affects the change of L x ( Figure 4 the length of OE in Fig. a), and L y ( Figure 4 the length of OV in Fig. a) remains unchanged, so this embodiment focuses on the analysis of L x . Figure 4 In, θ represents the incident angle of the left-edge light at x 1 , β represents the rotation angle of the RMS. In addition, ∠JRQ = π / 2 - θ = ∠IRS, ∠ISR = β, ∠RIS = π / 2 - α, so β = θ + α. In ΔIQS, |IS| = f 1 ·tanα

[20] , where f 1 represents the focal length of the half lens. According to the law of normal reflection, Figure 4 the length of KP in is calculated as follows:

[0045] KP = (H - d)·tan∠PRK = (H - d)·tan(α + 2θ)

[0046] where H = OS, representing the distance between the RMS and the top of the cabinet, d is the distance between the landing point of the incident light beam after the rotation of the RMS and the plane before rotation, and ∠PRK represents the angle between the line segment PR formed by points P and R and the line segment RK formed by points R and K. Then Figure 4 the length of OP in is calculated as follows:

[0047]

[0048] Similarly, it can be obtained that:

[0049]

[0050] Among them, ψ represents the abnormal reflection angle of the right-edge light at x m and is defined as follows:

[0051]

[0052] Among them, p represents the relative defocus parameter (p = 1 - v / f 1 ), where the distance between the output plane of the fiber array and the half mirror is v, and f 1 represents the focal length between the lens and the metasurface. λ represents the wavelength of the incident light, 1550 nm, and Ω = 4 μm represents the length of the RMS single supercell.

[0053] In addition, the highest bandwidth supply B m (T) of the defined topology Φ is as follows:

[0054] B m (T) = n × x × y × (m × n) × s Gbps

[0055] Among them, s represents the single-fiber rate; n represents the number of beam coverage areas, and x and y represent the number of cabinets falling into each coverage area in each row and column respectively; given the highest bandwidth supply B m (T) that just meets the predicted traffic bandwidth demand, the expected topological mode T[n, x, y] can be deduced inversely. In this embodiment, it is assumed that the area changes of all coverage areas are the same, that is, the RMS rotation angles corresponding to different coverage areas are consistent. Substituting β into the formula can obtain |OP|. If |OP| ≤ g, it indicates that this topological mode is feasible; otherwise, it is not advisable.

[0056] This embodiment gives a description of the DC topology reconstruction algorithm assisted by En-LSTM traffic prediction. For the next time slot τ (the value of which is determined by the sampling period of the traffic data), if the predicted traffic demand D τ > 0.95 × B m (T), it is considered that it has approached the highest network bandwidth that the current topological mode T can provide. Here, 0.95 × B m (T) represents that the bandwidth utilization rate reaches 95%, and this is used as the upper limit of the bandwidth utilization rate. By rotating the RMS and selectively increasing the values of n, x, and y, the expansion of the current topological scale T[n, x, y] to the expected topological mode T' that just serves the predicted traffic bandwidth demand is realized. If D τ ≤ 0.95 × B m(T), it is necessary to traverse the set of topological patterns T obtained by the exhaustive method and find the topological pattern T' with the highest bandwidth utilization rate. If T' is equivalent to T, there is no need to rotate the RMS; otherwise, the rotating RMS will cut T to T' to avoid unnecessary bandwidth waste. In this embodiment, the topological scale is regarded as a variable parameter, and the traffic generated by the prediction model is compared with the current actual traffic, and the topology with the highest bandwidth utilization rate is used as the best topological strategy at the current moment.

[0057] In this embodiment, to more comprehensively evaluate the traffic prediction accuracy of the designed En-LSTM model, in addition to performing a fitting analysis between the prediction and the actual traffic bandwidth requirements, the best topological pattern that exactly serves the actual traffic bandwidth requirements is defined as T". If T" does not match T', the relevant RMS rotation is a misleading operation and an error is generated. The specific process is shown in Algorithm 1.

[0058]

[0059] To illustrate the effectiveness of the present invention, in this embodiment, the transverse wave vector is set to k tL = 1 / 2·k 0 [cos(π / 2·(t - 1)+π / 4), sin(π / 2·(t - 1)+π / 4)], k tR = 1 / 2·k 0 [cos(t - 1), sin(t - 1)], substitute it into the formula to calculate the phase required for each period, and then determine the rotation angle and size of each nano fin. After obtaining the target phase distributions of LCP and RCP, obtain the corresponding nano fin sizes and rotation angles according to the target phase distributions. Figure 5 This is the light intensity distribution and light field distribution of this TMS chip under the normal incidence of linearly polarized light. The beams carrying different angular momenta are deflected to the set positions and focused, as Figure 5 , and the specific coordinates are x = -15.98μm, y = 16.13μm, x = 16.15μm, y = 16.10μm, x = -16.08μm, y = -16.10μm, and x = 16.03μm, y = -16.15μm, and the gap with the theoretical value is very small, thus verifying the good one-to-four communication ability of this TMS chip.

[0060] In this embodiment, a 42U standard rack is selected, and its length and width are both 0.6m. When the RMS is not rotated, the length and width of the communication coverage area are L x = 2.75m, L y= 1.68 m, the topological scale Φ = [4, 3, 5] of the maximum network bandwidth supply without rotation can be calculated, and at this time OP = 0 m; according to the rotation angle β required for different topological scales, for example, when Φ = [4, 3, 6], OE = 3.6 m, and the expected rotation angle β = 3.28° is solved. It should be noted that the maximum DC topological scale is Φ = [4, 3, 8], and the corresponding maximum rotation angle is 6.33°, because when Φ = [4, 3, 9], the length of OP is greater than 0.6 m and the original communication link is interrupted. In addition, we analyzed that the rotation delay required for the maximum angle of 6.33° is 5.3 ms, which is better than the performance of MEMS.

[0061] Next, this embodiment takes the real traffic data collected on the Abilene network in the United States as a sample for simulation experiment analysis. The data sampling period is 5 minutes, and there are 938 traffic data in total. Among them, the first 70% are used as training data samples, and the latter 30% are used as test set samples. As Figure 6 shown, this embodiment compares the prediction performances of LSTM and En-LSTM. In the test samples, the fitting of the predicted value curve of LSTM to the true value is not ideal, while En-LSTM is closer to the true value. At the same time, the mean absolute error (MAE) result of LSTM is 0.023, while the result of the En-LSTM model is 0.012 (its overall prediction error is smaller).

[0062] According to the change characteristics of the data samples, this embodiment tests the highest network bandwidth supply B m (Φ) of different topological scales. As Figure 7 shown by the dotted line, for example, the highest network bandwidth supply B m (Φ) of the topological scale Φ = [4, 3, 5] is 0.45 Gbps. The transmission rate s of each optical fiber can be calculated as s = 0.6×10 -4 Gbps. Based on this, this embodiment takes s as a known value and sequentially solves B m (Φ) under different topological scales and uses it as the input of the reconstruction algorithm. Figure 7 In 1 and 2 , t 1 and t 1 represent the current and the next time slots respectively, and the running topological scale of the t 2 time slot is Φ = [4, 3, 7]. However, as the adjacent time slot t τ in the future, the predicted traffic bandwidth demand D m > 0.95×B

[0063] Finally, for the two prediction models, in this embodiment, a comparison was made on 278 test samples in combination with the reconstruction algorithm. The traditional LSTM prediction model caused 25 misleading operations of RMS rotation, while the En-LSTM model proposed in the present invention only had 2 times, and its improvement rate was as high as 91%, which could better guide the DC topology reconstruction. In addition, this embodiment also analyzed the maximum topology reconstruction delay, including the execution time of the reconstruction algorithm and the maximum rotation angle delay, which was approximately equal to 5.8 ms, that is, the average topology reconstruction delay was only 5.8 ms - 5.3 ms = 0.5 ms, reaching the microsecond level, which was better than the performance of MEMS.

[0064] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A data center elastic wireless optical interconnection system based on RMS and TMS cascade, including a reflective metasurface and a transmissive metasurface, establishing a communication link between a source cabinet and multiple cabinets, characterized in that: The beam output from the first fiber array port on the top of the source cabinet is transmitted through the transmission metasurface to output N beams with different azimuths and the same output angle ξ, which are respectively transmitted to the corresponding second fiber arrays. Each beam output from the second fiber array is modulated by LCPC and focused by lens deflection, and then 2N different beams are generated through the reflection metasurface to cover the cabinet.

2. According to claim 1, a data center elastic wireless optical interconnection system based on RMS and TMS cascade, characterized in that: N depends on the chip size of the transmissive metasurface and the lattice constant of the on-chip unit structure.

3. According to claim 1 or 2, a data center elastic wireless optical interconnection system based on RMS and TMS cascade, characterized in that: The maximum value of N is 2×l m , l m is the maximum number of OAM channels that the chip of the transmissive metasurface can support, which is calculated as follows: Wherein, W represents the chip size of the transmissive metasurface, and σ represents the lattice constant of the on-chip unit structure of the chip of the transmissive metasurface.

4. According to claim 3, a data center elastic wireless optical interconnection system based on RMS and TMS cascade, characterized in that: The chip size of the transmissive metasurface is 45 μm×45 μm, the lattice constant σ of the on-chip unit structure of the chip of the transmissive metasurface is 950 nm, and the maximum value of N is 6.

5. According to claim 3, a data center elastic wireless optical interconnection system based on RMS and TMS cascade, characterized in that: The chip of the transmissive metasurface is an array composed of multiple units, each of which is composed of a SiO2 substrate and a Si nanofin, and the Si nanofin is located in the center of the SiO2 substrate.

6. According to claim 5, a data center elastic wireless optical interconnection system based on RMS and TMS cascade, characterized in that: The thickness of the Si nanofin is 1.2 μm.

7. According to claim 1, a data center elastic wireless optical interconnection system based on RMS and TMS cascade, characterized in that: The system also includes a traffic prediction model, which controls the rotation of the reflective metasurface according to the prediction results of the prediction model, so that the data center topology scale can change flexibly with the changing traffic at any time, avoiding the situation where excessive traffic in a static topology causes service denial or too little traffic causes redundant resource configuration.

Images