With the rapid development of the satellite Internet of Things (IoT), a large number of short-burst users are aggravating collisions and interference among users of the access network. To address this issue, several organizations and individuals have put forward some dynamic access schemes. However, for most of the proposed dynamic access schemes, it is necessary to know the exact number of future time slot access applications. At present, some load estimation schemes have been proposed in the literature, but the accuracy of these schemes is not high, and they can only achieve load estimation for current time slot.

To solve this issue, we propose a load estimation method based on the leading code state and parameter estimation. A load prediction method based on machine learning is also proposed. The load estimation method based on leading code status and parameter estimation analyzes the relationship between the probability of leading code in different states within the time slot of the satellite IoT and the number of requests for access to the current time slot. It gives the maximum likelihood parameter estimation expression and uses the maximum likelihood parameter estimation method to estimate the current time slot load. The load prediction method based on machine learning takes the estimated load value as its historical data, combining the Long and Short Term Memory (LSTM) network and the Auto Regressive Moving Average (ARMA) model to predict the future time slot load.

The simulation results show that the estimated error of the load estimation method based on leading code state and parameter estimation is less than 1%. The comprehensive error of the load prediction method based on load estimation results as historical machine learning data is about 6%.

The predicted error of the proposed load estimation and prediction method is within the acceptable range, thus offering accurate future slot access requests for dynamic access schemes.

Fiber Raman Amplifier (FRA) based on multi-pump technology has features of low noise, a wide gain bandwidth, and a controllable gain spectrum shape, which is regarded as an ideal optical relay amplifier for long-haul fiber optic transmission network systems. Intelligent optical amplifiers with adaptive controllable gain are required in dynamic fiber optic transmission network systems. This article introduces a gain control method for FRA based on neural network and numerical solutions of Raman power coupling equations.

First, the data set containing the signal gains, pump powers and wavelengths in the FRA is collected to train the neural network to establish an approximate mapping relationship between the signal gains and pump parameters. Subsequently, the trained neural network is utilized to determine the initial pump powers and wavelengths of the FRA based on the target gains of the signal. Finally, the pump powers are optimized by solving the numerical solutions of the Raman power coupling equations to improve the accuracy of the FRA output signal gains.

The paper investigates the effect of the flatness of signal gains in each group in the training dataset on the accuracy of FRA output signal gains. When the gain fluctuation of each group signal in the training data is less than 2 dB, the mean and variance of the Root Mean Square Error (RMSE) of the 1 000 sets of test signal gains output by the FRA are 0.230 and 0.010 dB, respectively. Additionally, the mean and variance of the maximum error of the gains are 0.462 and 0.044 dB, respectively.

The results indicate that the proposed method can achieve high-precision FRA gain control, offering a new idea and method for investigating intelligent optical amplifier gain adaptive control in dynamic fiber optic transmission networks.

The emergence of new technologies is driving the continuous improvement of fiber optic communication capacity, which poses new requirements for the working wavelength range of fiber amplifiers. Therefore, there is an urgent need for fiber amplifiers that can achieve ultra wideband amplification.

In order to achieve ultra wideband Raman Fiber Amplifier (RFA) with gain ranges covering Super-C and Super-L bands, a transmission model of RFA was theoretically established, and numerical solutions for the transmission model of backward RFA were introduced. The pump power and wavelength range were selected considering the actual situation. Six pump wavelengths and seven pump lasers were used to simulate an optimized set of pump wavelength configurations in G.652 fiber. Based on the configuration of these pump wavelengths, experimental tests were conducted on three types of fibers: G.652, G.654, and Large Effective Area Fiber (LEAF). The simulation and experimental results were compared and analyzed.

The results show that a net gain of 18.07 dB and a gain flatness of 1.54 dB were obtained through testing in G.652 fiber. A net gain of 13.9 dB and a gain flatness of 1.21 dB were obtained through testing in G.654 fiber. A net gain of 21.99 dB and a gain flatness of 1.80 dB were tested in LEAF. The difference between the simulation results and the test results is less than 0.2 dB, and the difference in gain flatness is less than 0.4 dB.

The comparison between simulation results and test results indicates that the numerical solution method used in this paper can accurately predict the gain spectrum of RFA. The pump wavelength configuration used in this article has been experimentally verified to have significant gain and good gain flatness among the three types of fibers, which has reference value for wavelength selection in commercial ultra wideband RFA.

At present, most of the proposed data center optical interconnection network architectures use optical networks for inter-rack communication among switches, while the intra-rack communication among servers is still realized by electrical Ethernet switches. Due to the large power and large number of electrical Ethernet switches, the existing data center optical interconnection networks have the problems of high energy consumption, low bandwidth provision and large packet delay.

In order to further improve the performance of data center optical interconnection network, this paper proposes an in-rack optical interconnection network architecture based on a rearrangeable non-blocking Clos network. In this architecture, servers in the rack are connected to each other through a three-stage Clos network. Non-blocking communication between servers in the rack can be achieved by the proposed multi-wavelength routing algorithm.

In this paper, OMNET+ + software is used for simulation verification. The simulation results show that the throughput of the proposed architecture is 20% higher than that of the traditional in-rack electrical switching under the same conditions. The average end-to-end delay is reduced by 90%, and the energy consumption is reduced by 67%.

It also demonstrates that the proposed architecture can improve the performance of bandwidth usage, packet delay and energy consumption in a data center optical interconnection network.

This paper aims to improve the multi-layer optical fiber wiring structure and the pressure resistance. An experimental research on pressure distribution optimization is conducted. A multi-dimensional optical cross-interconnection flexible optical backplane pressure distribution optimization structure and design method are also proposed.

This paper uses experimental analysis of different optical fiber crossing methods to study the influence of optical fiber arrangement, fiber intersection density and number of optical fiber overlap layers on optical fiber loss under the same pressure. It also proposes an ideal optical fiber wiring structure of double-layer backplane with optimized optical backplane pressure distribution.

The proposed ideal double-layer wiring structure adopts double-layer backplanes and optimized wiring of multiple layers of optical fibers in each backplane. The two flexible optical backplanes are stacked vertically, with the optical fiber ports evenly distributed along the four edges of the backplanes. Additionally, the optical fibers of the same layer are tightly arranged, and two layers of optical fibers overlap at the crossing points.By utilizing the layout of multi-layer optical fiber crossings and staggered overlays, an optimized distribution of multi-layer optical fiber cross-interconnect wiring can be achieved, enhancing the compressive resistance of the optical backplane and thereby improving its overall optical transmission and switching performance.

The proposed scheme features an optical backplane with a smaller volume, better compressive resistance, and improved optical signal transmission loss performance. It enables the realization of larger-area, high-density, and high-capacity flexible n×n optical backplanes with excellent scalability, making fiber optic wiring work simpler and more convenient.

With the rapid development of cloud computing, big data technology, and Data Center(DC), the scale and number of DC continue to expand, and the bandwidth resources and energy consumption of optical networks is also growing rapidly. It can be seen that the method can effectively reduce both bandwidth resources and energy consumption in the network has received widespread attention.

Considering the characteristics of both the network energy consuming components, the spectrum resource allocation, and the selection of different modulation formats, we employ both adaptive traffic grooming method and a spectrum layering method for network traffic grooming. A Distance-adaptive Stratification Grooming Method (DSGM) in elastic optical networks is also proposed. We can employ the adaptive selection of different modulation formats to divide spectrum resources into different spectrum layers for network traffic grooming. In this way, on the one hand, the distance adaptive modulation method can select different modulation formats based on the length of the connection request working path. On the other hand, based on the spectrum layered traffic grooming method, the idle spectrum resources of each fiber link are layered according to different rates of optical channels. The appropriate optical channels are selected for traffic grooming through optimizing the use of spectrum resources based on the remaining bandwidth resources of the fiber links. We also introduce three different traffic grooming methods for comparison, including Spectrum Stratification Grooming Method (SSGM), Traditional Traffic Grooming Method (TTGM), and Traditional Method Without Grooming (TMWG).

The simulation results show that the proposed approach significantly reduces the energy consumption, and the blocking probability. It also improves the network spectrum efficiency compared with SSGM, TTGM, and TMWG method.

We can realize the network resource optimization by using the stratification traffic grooming methods in elastic optical networks.

In the context of extensive data transmission and exchange operations in data centers, there is a high-performance demand for switching equipment. Switching equipment not only needs to exhibit excellent scalability but also must provide large bandwidth and low switching delay. Fast optical switches possess superior performance with extremely faster switching rates and access bandwidth than electrical switches. Faced with the ever-growing demand for data exchange in the present day, the structure of switches is evolving towards all-optical directions to achieve greater access rates and processing performance. In the study of fast optical switches, the scheduling algorithms focus on achieving a 100% throughput, with relatively less emphasis on ensuring Quality of Service (QoS).

This paper proposes an iterative scheduling algorithm for optical switches that ensures latency, based on the principles of matrix decomposition and traffic shaping. We combine iteration and round robin mechanisms, and propose the Iterative Round Robin (IRR) scheduling algorithm. Initially, the IRR algorithm maps the traffic rate matrix to a traffic matrix, and then calculates the service requirement matrix sequence based on the service requirements of the traffic matrix. Finally the scheduling matrix sequence is extracted from the service requirement matrix sequence.

Theoretical analysis validates that the IRR scheduling algorithm can provide port-based guaranteed latency and provides a method for calculating latency upper bounds under the given traffic conditions.

The IRR scheduling algorithm can be applied in optimizing the latency performance of optical switches and in designing optical switch systems with guaranteed latency.This contributes crucial support to the provision of deterministic QoS for the next generation all-optical networks.

Recent research efforts on Routing and Wavelength Assignment (RWA) for all optical networks are focused on Deep Reinforcement Learning (DRL) based algorithms. The DRL based RWA algorithms are mostly rely on the K Shortest Paths (KSP) routing to calculate candidate paths in advance, hence the DRL agent can choose possible actions from the precomputed paths. These KSP based models lack of flexibility and dynamicity, since they need to re-calculate the KSP for all the node pairs once the topology changes occur. To address this issue, this paper proposes an Adaptive and Efficient(ADE)-RWA algorithm based on DRL.

The key points and innovations of the ADE-RWA lie in that during the training process, the DRL agent takes actions in a step-by-step way instead of selecting from the precomputed K complete paths. Therefore, the routing strategies are dynamically adjustable in training even under the case of topology changes. It is because that the actions are open for the agent to take without concerning the limitations of the K fixed paths. Moreover, the ADE-RWA records the successfully assigned routes during the training in a LookUp Table (LUT). The algorithm turns to LUT checking for finding the available routes once the DRL training is converged, since at that time the LUT has acquired enough information for the RWA from the DRL training. The LUT based routing can effectively reduce the computational costs and improve the efficiency of RWA. In addition, the DRL training phase and LUT routing phase are real-time switchable. The algorithm turns to the DRL training phase when a link failure caused topology change occurs, and turns back to LUT checking when the model training is converged again.

Experimental results show that compared with KSP-First Fit(FF)and Deep Reinforcement Learning Framework for Routing, Modulation and Spectrum Assignment (DeepRMSA), the blocking probability of ADE-RWA is reduced by 36% and 30% respectively. When a link failure occurs, the algorithm can quickly adapt to the changes in network topology.

The proposed DRL based RWA framework ADE-RWA can achieve adaptive routing and wavelength allocation under dynamic network conditions with low computational cost.

A Loopback Interconnect Datacenter Optical Network (LIDON) is proposed for switching in datacenters based on commercial devices.

The architecture is mainly composed of two switching modules, the Rack Expansion Module (REM) and the Loopback Interconnect Module (LIM). The intra- and inter-Pod switching in datacenters is accomplished with those two modules operating collectively. By developing the new functions of diagnostic ports of a wavelength selective switch, LIDON boasts several features, such as high path diversity and robust tolerance. Moreover, the scalability of LIDON is guaranteed by the proposed scheme to support the flexible switching volume.

The feasibility of dividing intra- and inter-Pod signals in LIDON was demonstrated by the filtering performance experiments on the switching components. The simulations of the task completion time for different services measure the fault tolerance of LIDON.

The test and simulation results show that the proposed architecture has high path diversity and strong fault tolerance.

In recent years, Artificial Intelligence Generated Content(AIGC)has set off the artificial intelligence revolution. The network connection of the Intelligent Computing Center(ICC)has also developed in the direction of ultra-high bandwidth, intelligent lossless, and computing network convergence. Therefore, the optical network of the ICC needs to reduce the inter-card communication time in order to improve the efficiency of data access.

The paper addresses the networking architecture of optical networks for ICC scenarios to realize a lossless network with large bandwidth, low latency and high Central Processor Unit (CPU) efficiency, which can satisfy the demand of large model training and reasoning in ICC. This paper analyzes in detail the traffic distribution characteristics of the ICC and the communication flow characteristics under the AI large model training networking scenario. It also conducts in-depth research on the technologies such as Ethernet lossless network based on Remote Direct Memory Access(RDMA) technology and optoelectronic co-encapsulation. Finally it carries out the networking practice and latency test under the ICC scenario.

The RDMA over Converged Ethernet(RoCE)-based transport scheme proposed in this paper has the capabilities of priority-based flow control, displaying congestion notification, enhanced transport selection and data center bridge capability switching protocols, which can realize lossless transmission based on Ethernet protocols in data centers. The test results in this paper show that the transmission delay using the RoCE protocol is approximately stable at around 1 μs and significantly outperforms the Internet Wide Area RDMA Protocol(iWARP).

In this paper, based on the traffic characterization in the intelligent computing scenario, we have studied the key characteristics of the lossless Ethernet network in the ICC, and used the RDMA technology to realize the enhancement of the transmission efficiency of the optical switching network in the scenario of the ICC. We have also put forward a lossless Ethernet network scheme under the large model inference scenario of the ICC, and explored the feasible direction for the application of the RDMA technology in the intelligent computing scenario. The proposed scheme explores a feasible direction for the application of RDMA technology in the smart computing scenario.

The rapid development of artificial intelligence has put forward higher requirements for data center computing. As the core device of the data center optical interconnection, the design of optical module will also encounter big challenges. With the increasing speed of optical modules, the signal integrity problem has become a bottleneck that restricts the performance of optical modules. In order to design a high-speed optical module that can meet the speed requirements of optical interconnects in data centers, especially high performance computing application, it is necessary to implement optimal design on the ultra-high-speed link inside the optical module.

In this paper, the factors affecting the integrity of the signal in the optical module are simulated and optimized with a Quad Small Form-factor Pluggable Double Density (QSFP-DD) optical module as an example. The specific work is to theoretically analyze the parts of the channel that may deteriorate the signal integrity, such as vias and Ball Grid Array (BGA) solder balls, and discuss the optimization and improvement methods of these parts. In particular, this paper analyzes the optimization method for signal transmission in the frequency band above 50 GHz, so that the high-speed channel can realize the low-loss transmission for ultra-high-speed 4-Level Pulse Amplitude Modulation (PAM4) signal. In addition, the transmission performance of the whole high-speed channel is also studied and tested by the four-port vector network analyzer.

The results show that the design scheme can achieve the required transmission bandwidth without resonance. The simulation results of full channel show that the return loss of all channels is less than -15 dB and the insertion loss is less than 3 dB, and 224 Gbit/s PAM4 signal low-loss transmission can be achieved. The test results are basically consistent with the simulation results, which proves that the design scheme can meet the requirement of high speed optical interconnects in Datacenters.

The method of signal integrity design of optical modules proposed in this paper has important guiding significance for the design of ultra-high-speed optical modules in the future, and will also provide a reference for the design of other kinds of high-speed circuits.

To meet the requirements of Artificial Intelligence(AI)cluster networks for flexibility in resource allocation and resource utilization, Optical Burst Switching (OBS) has returned to people’s view. More items may be required to the optical module: firstly, a nanosecond optical path switching speed is required, and the receiver of the high-speed optical module is required to adapt to burst mode. Secondly, currently high-speed optical switches that reach the nanosecond level usually have the problem of high optical path insertion loss. Therefore, the link loss of high-speed optical modules is required to meet the Extended-Reach(ER) distance standard.

To meet this requirement, the article presents the assembly of an ER optical module using a fully commercial chip with a Thin Film Lithium Niobate (TFLN) and PhotoDetector (PD) solution. A testing system has been designed for this module, and the subsequent steps of the work are still in progress.

The experimental results show that an optical link budget of 20 dB is obtained, basically meeting the requirements of high-speed optical switching for optical module link budget. The OBS system is acceptable for optical modules with various particle size. We have preliminarily explored the design of three types of 100 Gbit/s lane optical modules: due to limitations in overall power consumption and volume, four wave mixing, and difficulties in coupling with multiple Avalanche PhotoDetectors(APD), the implementation of 4-way ER4 optical modules is much easier than 8-way ER8 modules. The ER2 module with two optical channels can adopt the coaxial process Electro-absorption Modulation Laser(TO-EML)+ small aperture APD scheme, which has the advantages of easy implementation, high cost-effectiveness, and mature supporting chips.

Therefore, 200 Gbit/s ER2 is the optimized choice for OBS systems

The Array Waveguide Grating Router (AWGR) constructed in the form of an N×N matrix has optical parallelism and wavelength routing capabilities. It can simultaneously transmit N signals on different channels, and has advantages of good scalability, low latency, and wide bandwidth. Combined with tunable harmonic light sources, it can achieve fast optical switching, which is one of the potential technical solutions for the next generation of optical switching data center networks. In order to solve the problems of high crosstalk, additional coupling loss, polarization sensitivity, and non-uniformity of existing AWGR that may affect practical applications, this paper studied AWGR with 4×4 channels and 12×12 channels respectively to further expand the scale of data centers and improve the data exchange speed.

The basic design parameters were calculated using simulation software, and the AWGR design process was analyzed and studied. The Beam Propagation Method (BPM) was used for simulation. At the same time, performance optimization is carried out by adding a conical waveguide tap structure at the connection between the flat waveguide and the strip waveguide, increasing the spacing between the input and output waveguides.

The simulation results show good performance parameters: 4×4 AWGR insertion loss of -0.714 dB, crosstalk of -35.556 dB, loss non-uniformity of 1.907 dB; 12×12 AWGR insertion loss of -0.294 dB, crosstalk of -36.019 dB, loss non-uniformity of 3.428 dB. The designed device chips are then fabricated and tested on the optical platform. The test results indicate: 4×4 AWGR insertion loss of -2.586 dB, crosstalk of -29.473 dB, loss non-uniformity of 1.921 dB; 12×12 AWGR insertion loss of -3.692 dB, crosstalk of -23.874 dB, loss non-uniformity of 3.873 dB.

This article investigates the performance optimization in areas such as crosstalk and non-uniformity of losses, accumulating experience for subsequent design iterations and further improving the performance parameters.

As a new optical communication technology, intelligent elastic optical network provides higher bandwidth and flexibility to meet the increasing demand of power communication. However, due to the uneven spectrum demand size and distribution of different communication requests, it may lead to fragmentation of spectrum resources in the network. These fragmented spectrum resources may result in the waste of spectrum resources and the degradation of network performance.

This paper introduces an evaluation method of fragmentation index, and takes it as a reference to judge whether to carry out spectrum defragmentation operations for running services.

The simulation results show that compared with the traditional algorithm, the bandwidth blocking rate can be reduced by 5% to 30%. It is also shown that the the bandwidth blocking rate reduction is more obvious when the traffic volume is larger.

The algorithm proposed in this paper comprehensively considers the proportion of spectrum fragments on a single fiber, the average frequency gap of services on the link, and the occupation weight of each time slot in each fiber core. This paper also evaluates the frequency gap occupation of the network from a multidimensional perspective, providing an efficient optimization scheme for resource allocation.

Compared to Electronic Packet Switching (EPS), Optical Circuit Switching (OCS) demonstrates advantages in latency, power consumption, cost, and stability. This study aims to explore feasible applications of OCS in the networking of training tasks by analyzing parallel partitioning strategies, collective communication requirements, traffic patterns, and current network architectures in large model pretraining, in order to fully leverage the benefits of OCS.

We propose a mechanism for network device redundancy protection using multiple small-port OCS devices, enabling rapid switching without interrupting training tasks in the event of Top-of-Rack (ToR) switch failures. Additionally, we advocate for the exclusive service of OCS to data parallelism, requiring configuration only at the start of the task.

We present several feasible opto-electronic networking architectures and specific configurations under different AllReduce algorithms, including joint optimization of collective communication algorithms and architectural design to achieve optimal bandwidth.

By adequately integrating the traffic models of training tasks, OCS can seamlessly blend into existing EPS network architectures and optimize the large model pretraining from multiple perspectives, including cost, low power consumption, reduced latency, and enhanced stability.