ObjectiveConventional millimeter-wave mobile forward transmission systems usually utilize electronic devices to realize signal generation and processing, with their frequency bands and bandwidths limited by the “electronic bottleneck”. Microwave photonics technologyhas the advantages oflow loss, high frequency, large bandwidth and reconfigurable, which can effectively break through the “electronic bottleneck”. It alsoenable the generationand long-distance transmissionof high-speed wideband signals. However, current photogenerated millimeter-wave methods based on Electro-Optical Modulatier (EOM) are often complex in structureand difficult to control.MethodsTo address this problem, a programmable optical frequency multiplication and optical frequency comb generation method is proposed in this paper, with all optical functions integrated in a Dual-Polarization In-phase and Quadrature Modulator (DPIQM). Utilizing the automatic bias control technique based on digital chaotic dither and harmonic correlation detection, the DPIQM can be controlledat different bias states, allowing the generation of Optical Frequency Combs (OFC) with highflatness and optical frequencymultiplication signals with highsuppression ratio.ResultsA prototype is designed and experimentally demonstrated based on the proposed method in this paper. A 9-line OFC with a Free Spectral Range (FSR) of 8 GHz and a power flatness of up to 0.86 dB was experimentally generated. Optical frequency multiplication signals of 1st to 4th order with an FSR of 2 to 4 GHz and a harmonic rejection ratio of more than 13 dB were also generated. In addition, the generated OFC and optical frequency multiplication signals were also experimentally proven to haveexcellent phase noise characteristics. The phase noise of the electrical signal obtained by beating the 4th order optical frequency multiplication signal is lower than-115 dBc/Hz@1 MHz.ConclusionThe programmable optical frequency multiplication and OFC generation method provided in this paperis integrated in function, compact in structure, flexible in control, convenient in adjustment, and can realizemillimeter-wave signal with up to 8 folds of frequency multiplication. Ithasimportant practicalsignificanceto expand the bandwidth and rate of future multi-band millimeter-wave communication systems.
ObjectiveWith the development of deep learning large models, the demand for computing power is growing exponentially, requiring intelligent computing networks to have high bandwidth, low latency, and high reliability. Optical switching has been introduced into the artificial intelligence data center network due to its characteristics of transparent speed, no photoelectric conversion, and low power consumption. This paper proposes a hybrid optical/electrical network architecture for distributed deep learning based on Array Waveguide Grating Router (AWGR) and tunable lasers. A network resource priority scheduling scheme based on flexible bandwidth allocation is also designed.MethodsA network simulation model is built using OMNeT++ software, and the priority scheduling scheme is validated in the simulation model. The network performance of the architecture in distributed training traffic mode is evaluated.ResultsThe simulation results show that the architecture achieves end-to-end error free and zero packet loss communication in the neighbor traffic mode represented by Ring-AllReduce. The average end-to-end latency is reduced by 33.37% compared to the Fat-Tree architecture, and the communication latency between edge layer switches is reduced by 89.12% compared to the Fat-Tree architecture. The delay fluctuation between edge layer switches can be maintained at no more than 5.42% under different loads. The paper analyzes and compares the network construction cost and power consumption of this architecture and the Fat-Tree architecture in two different scales of networks. The analysis results show that the network construction cost of this architecture is 67.52% lower than that of Fat-Tree, and the total power consumption is 66.69% lower than that of Fat-Tree.ConclusionCompared with the Fat-Tree architecture, this architecture automatically reconstructs the optical switching network by introducing a nanosecond-level fast optical switching mechanism to adapt to the periodic changes in network traffic generated by distributed training. It also significantly reduces the delay of cross-group communication with the same card number by reducing the number of hops experienced during the communication process, resulting in a 1.5-fold acceleration of the network when executing Ring-AllReduce tasks.
ObjectiveThis paper proposes a low-rate signal transmission protocol based on Field-Programmable Gate Array (FPGA), aiming to solve the protocol blocking problem caused by bandwidth differences in power optical fiber communication. It also provides a reliable transmitting and receiving handshake mechanism to support stable signal transmission in an environment with extremely different bandwidths.MethodsA transmission protocol with a bandwidth of 2 kbit/s at the transmitting end and a bandwidth of 625 Mbit/s at the receiving end is designed. The experimental system includes five modules: data packet reception, data framing, forward error correction, decoding, and receiving verification.ResultsExperimental results show that under the extreme bandwidth different conditions-specifically, 2 kbit/s transmission and 625 Mbit/s reception over 425 km of fiber-the system achieves an Signal Noise Ratio (SNR) of 8.58 dB and maintains a forward-error correction bit error rate of 5.17×10-5. The low-rate signal is successfully framed and transmitted to the receiving end. After transmission and Analog-to-Digital Converter (ADC) sampling at the receiving end, the frame is accurately dismantled and the received signal is verified, thereby confirming the reliability and applicability of the proposed method for power fiber communication networks.ConclusionThe low-rate signal protocol handshake method proposed in this paper, implemented on an FPGA, achieves effective error correction as well as transmission and reception verification. It runs stably in a simulation environment and resolves the protocol blocking problem caused by bandwidth asymmetry. This approach provides an efficient solution for the reliable transmission of low-rate signals, which has practical application value.
ObjectiveThe goal of the future 6th Generation Moblie Communication Technology (6G) wireless communication system is to achieve a peak data rate of up to terabits per second and reduce the delay to sub-millisecond level. However, the free space path loss of Terahertz (THz) communication is large, the diffraction and diffraction capabilities are weak, and the coverage is relatively limited. In order to deal with the large space path loss to support user mobility and serve multiple users, THz beam steering technology is essential.MethodsIn this paper, three microstrip periodic leaky-wave antenna arrays with monolithic integrated Beamforming Network (BFN) on indium phosphide substrate are designed. The purpose is to utilize these designs in photon-assisted two-dimensional terahertz beam steering chips.ResultsThe designed leaky-wave antenna units employ a mirror-asymmetric structure to suppress the open stopband effect. A single leaky-wave antenna is cascaded by 20 unit structures. It has a continuous beam scanning capability of-55~50° in the operating frequency range of 230~330 GHz. For the chip integrating three leaky-wave antenna arrays with a 2×3 coherently radiating periodic structures BFN, the total beam elevation scanning range reaches 105°, while the total azimuth scanning range attains 58° at a delay difference of 1.79 ps.ConclusionThe chip integrates three microstrip periodic leaky-wave antenna arrays with a single-layer 2×3 coherent radiation periodic structures beamforming network, enabling two-dimensional beam steering and making it well-suited for photon-assisted THz beam steering applications.
With the rapid development of sensing technology, Fiber Bragg Grating (FBG) strain sensors have gradually become practical and commercialized. Compared with traditional sensors, FBG strain sensors are widely used in fields such as highways, buildings, bridges, dams, mining areas, railways, oil or gas reservoirs. They have gradually become one of the important sensors in the detection field due to their advantages of small size, high sensitivity, corrosion resistance, and strong anti-interference ability. At present, FBG strain sensors are developing rapidly and extending into many new directions. This article focuses on FBG strain sensors and analyzes and discusses the progress of their extended applications in displacement, acceleration, humidity, corrosion, and ultrasound directions. It details the theoretical principles of each extended application, provides a comparative analysis of their performance, and offers a perspective on future extended applications of FBG strain sensing technology.
ObjectiveDue to the multi-dimensional modulation capability, coherent technology is emerging as a core evolutionary direction for high-speed intelligent computing optical interconnects. However, the high cost and power consumption remain key bottlenecks restricting its deployment in short-reach scenarios. To effectively reduce its implementation cost and power consumption while maintaining the advantages of high-capacity and high-performance, simplified coherent technologies for Artificial Intelligence (AI) supernode interconnects have become a hot research topic.MethodsTo tackle the cost and power bottlenecks inherent in conventional coherent architectures, this article systematically reviews existing simplified coherent solutions from three perspectives: low-cost phase/frequency-locking techniques, hardware-efficient Digital Signal Processing (DSP) algorithms, and real-time implementations. In terms of low-cost phase/frequency-locking techniques, a frequency offset locking scheme based on pre-scanning combined with Fast Fourier Transform (FFT) and Chirp Z-Transform (CZT), as well as a digital heterodyne optical phase-locked loop scheme combining coarse frequency offset acquisition with fine phase tracking, are proposed to avoid using expensive narrow-linewidth lasers. As for hardware-efficient DSP algorithms, a single-pilot-tone-based polarization demultiplexing method and a novel baud-rate 3Sign-Timing Error Detector (TED) scheme based on sign operations are proposed to reduce the hardware requirements of the equalization and clock recovery modules. Besides, a low-complexity real-time coherent system is verified based on 32 Gbaud dual-polarization Quadrature Phase Shift Keying (QPSK) signals.ResultsExperimental results demonstrate that the proposed phase/frequency-locking techniques can effectively ensure frequency-locking accuracy and frequency stability, with frequency fluctuation suppressed within 200 kHz. In terms of simplified equalization, the proposed scheme not only reduces the complexity of the equalization module but also improves the polarization tracking speed to 10 Mrad/s. As for the clock recovery, the proposed 3Sign-TED scheme can effectively avoid any multiplication operations, exhibiting extremely low hardware complexity. Furthermore, based on the 32 Gbaud dual-polarization QPSK real-time coherent platform, a low-complexity real-time coherent system with a net data rate exceeding 100 Gbit/s is successfully demonstrated.ConclusionThe above technical breakthroughs provide a reference for future high-speed optical interconnects in AI data centers.
ObjectiveWith the requirements of increased capacity in optical communication systems and the developments of few-mode-fiber-based space-division-multiplexing systems, the Few-Mode Erbium-Doped Fiber amplifier (FM-EDFA) utilized as a repeater has attracted more attention. However, the Differential Modal Gain (DMG) limits the transmission distance of Mode Division Multiplexing (MDM) systems, and degrades the system robustness.MethodsTo address these issues, this paper proposes a layered-distributed refractive index Few-Mode Erbium-Doped Fiber (FM-EDF) based on a ring-core structure. The fiber features optimized layer width and refractive index distribution, and incorporates a low-refractive-index groove. This design enables gain equalization with high gain and low DMG in a four-mode FM-EDFA (supporting LP01, LP11, LP21, and LP31) under cladding-pumping conditions.ResultsThe results demonstrate that introducing a layered-distributed refractive index structure into the high-index ring-core enables a DMG of 0.2 dB. Additionally, over the wavelength range of 1 530~1 565 nm, the proposed design achieves a gain exceeding 19 dB, a DMG below 0.23 dB, a noise figure lower than 5.96 dB, and a gain flatness of less than 0.9 dB. Then, investigations on the transmission characteristics are executed employing 1 000 km transmission systems with the repeatered distance of 25, 50, 100 and 125 km respectively. The results show that around 20 dBm output power and 4.05 dB DMG could be achieved over 1 000 km transmission with an optimal repeatered distance of 50 km.ConclusionThe scheme that introducing layered-distributed refractive index when designing the FM-EDF could effectively decrease the DMG, contributing to gain equalizations and all-optical amplifying of cladding-pumped FM-EDFA. It can also enhance the capacity of mode division multiplexing system benefitting from superior gain characteristics in long-haul transmissions.
ObjectiveTo address the issue that traditional evaluation methods struggle to quantitatively assess Artificial Intelligence (AI)-driven autonomous capabilities in the intelligent evolution of optical networks, this paper proposes a multidimensional intelligent capability evaluation method. This method aims to resolve core problems such as the insufficient adaptability of static evaluation mechanisms in dynamic service scenarios and poor cross-vendor compatibility.MethodsThe paper designs an optical network Capability Maturity Index (CMI) by constructing a three-level indicator system of 'infrastructure–algorithm–application', and integrates a dynamic weight adjustment mechanism to achieve precise quantification. Firstly, we establish an evaluation model that includes 9 core indicators such as optical layer programmability and fault self-healing rate, and innovatively introduce the dimension of knowledge generalization ability. Then, an online weight adjustment algorithm based on Long Short Term Momery (LSTM) is proposed. It is combined with a sliding window mechanism to achieve environment adaptation. Finally, a distributed evaluation system was developed and validated through a simulation platform (OMNeT++ simulating a 12 node topology) and a multi vendor network in the laboratory. Typical faults such as fiber breakage and wavelength conflicts were considered for comparative analysis.ResultsThe experiment showed that the dynamic weighting mechanism can reduce the volatility of evaluation results by 42%, and the fault localization time (2.73±0.41 s) was reduced by 44.5% compared to traditional methods. The optimization of resource allocation latency is 38.6 ms, which is reduced by 51%. The fault missed detection rate in a cross vendor environment is only 0.8%, which is better than the static weight method (5.7%) and the traditional Telecommunications Management Network (TMN) method (23.6%). The stability of CMI index in sudden traffic scenarios reaches±0.11, meeting the threshold requirement of ±0.15.ConclusionThis method significantly improves the accuracy and environmental adaptability of intelligent evaluation of optical networks through hierarchical quantization and dynamic weighting mechanism. It provides a standardized evaluation tool for device selection and technological evolution. In the future, further integration of digital twin technology is needed to enhance the ability to emulate the physical layer disturbances.
The rapid advancement of Artificial Intelligence (AI) technologies has led to an exponential increase in computing power demand, resulting in a sharp rise in the power consumption of intelligent computing centers. Focusing on the needs of intelligent computing centers, this paper focuses on the low-power optical interconnect technologies. It reviews the latest advances along two major paths including pluggable optics and chip-level solutions. It investigates industrial characteristics and standardization trends, analyzes domestic and international gaps, and proposes three recommendations to accelerate the development of low-power optical interconnects. Overall, preliminary progress has been made in low-energy optical interconnects, with growing research interest in related technologies and standardization, as well as accelerated exploration and practices toward mass production. It is essential for China to explore a development path that aligns with its national conditions, based on the characteristics of its own industrial foundation.
ObjectiveThis article proposes an innovative solution to the difficulty of power monitoring during the propagation process in coherent optical communication systems. It deals with the difficulty of real-time monitoring of power fluctuations caused by signal attenuation and abnormal losses, as well as the cost and system complexity issues introduced by additional monitoring devices.MethodsSpecifically, we have built a 5×50 km single polarization 16 Quadrature Amplitude Modulation (QAM) system with transmission power ranging from 1 to 6 dBm. The abnormal losses of 3, 5 and 7 dB are introduced at different positions in the link. At the receiving end, using the received signal and Digital Signal Processing (DSP) methods, the Power Profile Estimation (PPE) and abnormal loss localization functions in the fiber optic link were implemented based on Correlation Method (CM) and Minimum Mean Square Error (MMSE) method.ResultsThe simulation results show that both CM and MMSE methods can effectively identify and locate the location of abnormal losses in the link. However, CM can only roughly depict the curve of power variation trend and cannot accurately provide the actual power value. In contrast, the MMSE rule can accurately locate abnormal losses and estimate the true power in the link more accurately. In addition, as the transmission power decreases, the performance of PPE will gradually decrease. Finally, we achieved power curve estimation and abnormal loss localization in a 5×50 km, single polarization 16QAM system.ConclusionThis article implements power curve estimation and abnormal loss localization based on CM and MMSE. It also tests the performance impact of key system parameters including transmission power, abnormal loss size and location on PPE, which demonstrates the feasibility of the PPE method.
ObjectiveThis paper deals with the challenge of rising transmission loads and latency sensitivity in the new power grid, necessitating faster communication network speeds.MethodsA low-latency Synchronous Digital Hierarchy (SDH) frame transmission scheme is proposed by Frequency Shift Keying (FSK) modulation. At the transmitter of the grid, the pointer signal is converted from amplitude modulation to frequency modulation, and then from Delta-Sigma Modulation (DSM) to a small-amplitude digital signal, which is superimposed with the payload digital signal for synchronous transmission. At the receiver side, the weak pointer signal is obtained using oversampling quadrature detection while directly receiving the payload signal.ResultsUnder the superposition scheme based on FSK modulation, the pointer signal is detected at the receiver when the Bit Error Rate (BER) is lower than the threshold of the Forward Error Correction (FEC) technique. The minimum amplitude ratio of the required digital signal and the service signal is 0.06. At this time, the signal-to-noise ratio to satisfy the BER threshold is about 9 dB, and the power cost to the service signal is only 0.1 dB, which makes the service signal at the receiving end also accurately recognized.ConclusionThis paper proposed a pointer signal co-transmission scheme based on FSK modulation. By converting the pointer signal into a frequency modulated signal and then converting it into a digital signal by DSM, it is directly superimposed on the traffic signal for co-transmission. This scheme reduces the feedback latency and improves the control efficiency of the electric power communication network.
Hollow-Core Optical Fiber (HCF) has emerged as a prominent research hotspot, drawing considerable attention from researchers worldwide due to its exceptional properties, including ultra-low loss, ultra-low nonlinearity, and ultra-low delay. Domestic telecommunication operators and institutions are actively deploying projects to advance the development of hollow-core fiber-based communication systems. However, inherent mechanistic challenges associated with HCF, such as precise length measurement, long-term operational durability, and high-stability performance under real-world environmental conditions, raise critical concerns. These issues pose substantial barriers to the practical implementation of HCF in large-scale communication infrastructure. To harness the transformative potential of HCF technology for communication systems, efforts must be made to address these challenges systematically. This endeavor is anticipated to be a long process, requiring foundational advancements in specific application scenarios, such as short-reach interconnects within data centers.
ObjectiveOptical Circuit Switching (OCS) systems play a pivotal role in optical communication networks. However, their performance is significantly limited by coupling inefficiency. This paper aims to investigate and optimize an aspherical lens-based coupling scheme to enhance the coupling efficiency of the OCS system, mitigate the effects of geometric aberrations, and improve the system stability and overall performance.MethodsThis study examines the impact of various coupling mismatches on system insertion loss, with a focus on mode-spot and axial mismatches. Leveraging the unique light-focusing and aberration-correcting properties of aspherical lenses, a new aspherical lens design was developed and optimized for the OCS system. This approach seeks to improve the system coupling efficiency while minimizing the influence of geometric aberrations.ResultsThe optimized aspherical lens design significantly enhances the coupling performance of the OCS system, increasing coupling efficiency from 79.57% to 99.92%, approaching the theoretical maximum. The adverse effects of mode-spot mismatch and axial mismatch are notably reduced, and geometric aberrations are minimized. These improvements contribute to better optical path uniformity and stability, thereby supporting the reliable operation of the OCS system.ConclusionThe optimized aspherical lens design effectively addresses the insertion loss caused by mode-spot and axial mismatches, substantially improving the coupling efficiency and optical path stability of the OCS system.
ObjectiveChina ranks among the world leaders in the number of hydraulic structures, such as dams. The use of optical fibre sensing technology for hydraulic structures leakage monitoring technology is one of the hot directions in recent years. The purpose is to enhance fiber-optic-based monitoring technology for identifying leakage locations in seepage walls.MethodsA physical model of concrete leakage monitoring was built using Brillouin Optical Time Domain Analysis (BOTDA) sensing fibre optic active heating technology. The study explored the fibre optic temperature sensitivity law and leakage identification effect by changing the fibre optic heating power, fibre optic deployment spacing, and leakage rate.ResultsThe experimental results show that, the technology of active heating has a high level of sensitivity in detecting concrete leakage, particularly when the heating power is high. The ability of the fibre to sense leakage decreases as the distance between the leakage channel and the heatable sensing fibre increases. The relationship between the leakage rate and the fibre-optic monitoring of temperature change is inversely proportional. However, the degree of influence is also related to the fibre-optic heating power and the distance between the leakage channels.ConclusionThe BOTDA sensing fibre-optic active heating technology is a viable method for monitoring concrete leakage. However, cost, safety and environmental impact need to be considered in practical applications.
This study aims to analyze the standardization process of Non-Terrestrial Network (NTN) technology for 5th Generation Mobile Communication Technology (5G) and explore the evolution of the Release 17 (R17) and Release 18 (R18) of the 3rd Generation Partnership Project (3GPP). First, the architectural characteristics of NTN are systematically introduced. Then, the protocol enhancements of the R17 and R18 are summarized by comparative methods. The feasibility of some functional enhancements of the R17 is analyzed in combination with relevant simulation data, as well as the functional enhancements of the R18. Finally, based on a summary of existing standardization progress, this paper analyzes key development directions such as the deep integration of NTN and Terrestrial Network (TN), AI-assisted network management, and the continuous evolution of Internet of Things (IoT)-NTN, providing a reference for further research and industrial practice.
ObjectiveThe Deformable Mirror is a key component of Adaptive Optical (AO) systems, which is responsible for real-time correction of wavefront aberrations to enhance imaging quality and precision. However, systematic approaches for determining two key parameters of deformable mirrors-the coupling value and the Gaussian exponent-are currently lacking. This study aims to investigate optimal methods for selecting these parameters.MethodsUsing a 37-unit deformable mirror as a model, we established a comprehensive mathematical framework encompassing the deformable mirror system and wavefront reconstruction methodology. Through numerical simulations, we derived relationships among the residual wavefront Root Mean Square (RMS) value, the condition number of the slope response matrix, and the maximum actuator deformation with respect to the coupling value and Gaussian exponent. The simulations investigated the impacts of these parameters on wavefront correction accuracy, system stability, and structural rigidity. Based on a multifaceted analysis integrating these aspects, we identified the optimal coupling value and Gaussian exponent. Their correction efficacy are also validated.ResultsFor the 37-unit deformable mirror, the optimal coupling value was determined to be within the range of 0.15~0.20, and the optimal Gaussian exponent was established as 2.0. The wavefront correction capability of the deformable mirror under these optimal parameters was verified, thereby substantiating the effectiveness of the proposed parameter selection methodology.ConclusionThe numerical simulation approach presented in this study effectively facilitated the analysis and selection of the optimal coupling value and Gaussian exponent for a 37 unit deformable mirror. This method is scalable to deformable mirrors with varying numbers of units, offering a novel reference for the selection and parameter optimization of deformable mirrors.
ObjectiveWith the rapid growth of large-scale model parameters, a single data center is increasingly constrained by computing capacity, power supply, physical space, and network bandwidth. Interconnection among intelligent computing centers has therefore become essential for cross-domain parallel training. However, in cross–data center scenarios, heterogeneous communication flows generated by data, tensor, and pipeline parallelism are closely coupled with routing, spectrum allocation, and link occupation in the underlying optical network, which may increase communication latency, aggravate local congestion, and lead to imbalanced computing resource utilization.MethodsTo address these issues, this paper develops a joint computing-network optimization model for distributed large-model training across interconnected intelligent computing centers. It analyzes the communication mechanisms of data, tensor, and pipeline parallelism, and builds a computation–communication time model covering stage computation cost, inter-layer communication cost, and data-parallel global synchronization cost. Model stage partitioning, node selection, computing and storage constraints, together with optical-network routing and spectrum allocation constraints, are then formulated as a unified combinatorial optimization problem. The scheduling process is further modeled as a Markov decision process, and a Proximal Policy Optimization (PPO)-based intelligent scheduling algorithm is proposed. Within this framework, an Actor-Critic structure generates stage partitioning and node-placement decisions, while feasibility checking and latency feedback are incorporated to support iterative policy refinement.ResultsSimulation results show that, compared with static, nearest-node, resource-aware, and greedy strategies, the proposed PPO-based method can better match training-task requirements with the underlying optical-network state under multi-task and high-load conditions. It reduces request blocking probability and average training iteration latency while maintaining high computing-resource utilization and low variance of node computing load.ConclusionThe proposed method realizes coordinated optimization between distributed large-model training tasks and optical-network resources, providing a feasible solution for efficient training scheduling and computing-network resource management in cross–data center intelligent computing interconnection scenarios.
ObjectiveThis study aims to address the challenge of identifying blockages in the coal gangue slurry filling pipeline by proposing a novel method based on Distributed Acoustic Sensing (DAS) and analyzing the factors influencing the accuracy of this identification.MethodsBased on the principles of flow-induced vibration and DAS monitoring, a test platform was developed to emulate the blockages in a filling pipeline. By using a valve to simulate blockage locations, we analyzed amplitude changes before and after the simulated blockage, even without physical blockages. Additionally, we determined the optimal angle for adhering to the optical fiber. The positioning error for blockage detection using DAS was quantified at 0.6 m through tapping and water bath heating methodologies.ResultsThe experimental results demonstrate the effective identification of blockage locations within the filling pipeline based on the proposed method. The optimal angle for the optical fiber layout during blockage incidents was determined to be 60°, providing the most effective results and offering a solid theoretical support for practical engineering applications.ConclusionThe method presented in this study successfully identifies the blockage locations within the filling pipeline, enabling timely interventions when blockages occur. By analyzing the factors influencing pipeline blockage identification, a reasonable optical fiber layout method and angle are proposed. This provides a new monitoring approach and technical support for the safe operation of gangue slurry filling pipelines, offering practical value and potential for further promotion.
ObjectiveAs Artificial Intelligence (AI) accelerates, the parameter scale of foundation models continues to expand. The demand for computing resources in training and inference links continues to increase, catalyzing the deployment of massive AI computing infrastructures. In intelligent compute fabrics, inter-node communication overhead is the chief bottleneck stifling compute utilization. Consequently, profiling network performance with discrete-event simulators before build-out is paramount for optimizing design and slashing these overheads. However, the amount of communication traffic is huge and has a time-space dependence. Traditional full simulation faces huge state overhead, and there is a contradiction between network scalability and simulation efficiency.MethodsFirstly, we analyzed the communication characteristics in the distributed training scenario of large models. The set communication primitives of the dense and sparse models of the DeepSeek architecture are modeled as point-to-point traffic with spatio-temporal dependencies, and the traffic characteristics are analyzed. Then, based on the periodicity of the obtained traffic and the convergence of the cluster network during the training of the large model, a steady-state extrapolation acceleration algorithm is proposed. The OMNeT++ platform is used to capture the transient changes of the network during the simulation process, and the packet-level simulation of the communication traffic is carried out at the initial stage of the simulation. After the network enters the steady state, the discrete packet-level simulation is abstracted as a flow level. The flow rate is analyzed according to the principle of maximum and minimum fairness, and the simulation process of the remaining traffic is mathematically extrapolated.ResultsThe above simulation acceleration effect is verified on three network scales. Experiments show that this method achieves a maximum of 4.7 times acceleration of the total simulation time while retaining the underlying congestion characteristics of the network. It can also reduce the memory usage of the card scale simulation by 30%. Subsequently, a small-scale verification was performed on a 100,000-card network. Under the premise that the network characteristics obtained by simulation are close to the benchmark, the simulation time is reduced by 73%, and the memory usage is reduced by 31%.ConclusionIn this paper, collective communication is modeled as fine-grained point-to-point traffic, which provides data support for the evaluation of intelligent computing networks. The proposed steady-state extrapolation acceleration algorithm effectively improves the simulation efficiency of large-scale intelligent computing networks and provides a reference for the performance evaluation of larger-scale intelligent computing networks.
The development of generative artificial intelligence and Large Language Models (LLM) has driven the continued scaling of distributed training. Therefore, communication overhead has emerged as a major performance bottleneck, shifting the system constraints from computation to network connectivity. Traditional electrical interconnects are increasingly limited by band-width-distance tradeoffs, energy efficiency, and scalability, making them insufficient for large-scale synchronized data exchange. In contrast, optical interconnection technologies, with their high bandwidth density, low latency, and topological reconfigurability, provide a promising solution to these challenges. Accordingly, this paper systematically reviews the key architectures and core technologies of optical interconnect networks for intelligent computing centers and discusses future development trends.
