arXiv subjects
Publications and source records attributed to Weisheng Hu.
We present the first LLM-powered multi-agent system for autonomous fault management across millions of optical links in production AIDCs. Refined via SFT and continuous memory evolution, it achieves 97.7% F1 and over 60% fault-incident reduction, outperforming SOTA LLMs on a ten-week field data evaluation.
The rapid scaling of large language models (LLMs) exacerbates communication bottlenecks in AI data centers (AIDCs). To overcome this, optical circuit switches (OCS) are increasingly adopted for their superior bandwidth capacity and energy efficiency. However, their reconfiguration overhead precludes intra-iteration topology update, necessitating a priori engineering of a static topology to absorb time-varying LLM traffic. Existing methods engineer these topologies based on traffic matrices. However, this representation obscures the bursty concurrent bandwidth demands dictated by parallelization strategies and fails to account for the independent channels required for concurrent communication. To address this, we propose DELTA, an efficient logical topology optimization framework for AIDCs that leverages the computation-communication directed acyclic graph (DAG) to encode time-varying traffic patterns into a Mixed-Integer Linear Programming (MILP) model, while exploiting the temporal slack of non-critical tasks to save optical ports without penalizing iteration makespan. By pioneering a variable-length time interval formulation, DELTA significantly reduces the solution space compared to the fixed-time-step formulation. To scale to thousand-GPU clusters, we design a dual-track acceleration strategy that combines search space pruning (reducing complexity from quadratic to linear) with heuristic hot-starting. Evaluations on large-scale LLM workloads show that DELTA reduces communication time by up to 17.5% compared to state-of-the-art traffic-matrix-based baselines. Furthermore, the framework reduces optical port consumption by at least 20%; dynamically reallocating these surplus ports to bandwidth-bottlenecked workloads reduces their performance gap relative to ideal non-blocking electrical networks by up to 26.1%, ultimately enabling most workloads to achieve near-ideal performance.
In this paper, the quasi-constant modulus (QCM) property is analyzed and leveraged in the design of nonlinearity-tolerant four-dimensional (4D) modulation formats. Accordingly, we propose a family of QCM-based quadrature amplitude modulation (QCM-QAM) constellations with high spectral efficiencies (SEs) of 9, 11, and 13 bit/4D-sym, respectively. The quasi-constant modulus design theoretically enhances tolerance to fiber nonlinearities. Meanwhile, QCM-QAM is evaluated in an unrepeatered wavelength-division multiplexing (WDM) system over both standard single-mode fiber (SSMF) and non-zero dispersion-shifted fiber (NZDSF). Across all SEs, QCM-QAM demonstrates robust nonlinear tolerance in both SSMF and NZDSF. This is evidenced by a consistent shift of the optimal launch power toward higher values and a significant improvement in effective signal-to-noise ratio (SNR). QCM-QAM also delivers generalized mutual information (GMI) gains of 0.22, 0.09, and 0.21 bit/4D-sym in SSMF, and 0.24, 0.10, and 0.22 bit/4D-sym, in NZDSF at the optimal transmission power, corresponding to the SEs of 9, 11, and 13 bit/4D-sym. Furthermore, QCM-QAM achieves transmission reach extensions of 1.6%, 0.9%, and 1.7% in SSMF, and 1.7%, 1.5%, and 1.8% in NZDSF, respectively, for the three SE levels.
Communication is pivotal in LLM training, and a thorough analysis of the communication efficiency of AI data center (AIDC) network is essential for guiding the design of these capital-intensive clusters. However, conventional metrics are inadequate for such analysis, as they do not directly link network activity to computational progress and lack granularity to diagnose the impact of different network design patterns. To address this, we introduce a metric framework, the Switching Efficiency Framework, whose core metric - Switching Efficiency ($η$) - quantifies computationally effective data throughput per unit switching capacity. We further decompose $η$ into three factors - Data, Routing Efficiency, and Port Utilization to facilitate analysis of distinct communication bottlenecks. Using this metric framework, we demonstrate how the symmetric, distributed switching of 3D-Torus and the centralized, hierarchical switching of Rail-Optimized architecture align with sparse or imbalanced LLM training traffic, and show that All-to-All traffic from Mixture-of-Experts models severely degrades their port utilization and routing efficiency. Our analysis also demonstrates how key design choices - such as adjusting switching resource allocation, expanding server size, adopting in-network computing, and multi-plane design - positively influence distinct facets of communication efficiency. Ultimately, the Switching Efficiency Framework provides an analytical tool for analyzing efficiency bottlenecks, thereby informing the design of future-generation AIDC networks.
We propose a communication-bound-aware cross-domain resource assignment framework for pipeline-parallel distributed training over multi-datacenter optical networks, which lowers iteration time by 31.25% and reduces 13.20% blocking requests compared to baselines.
Fast and accurate waveform simulation is critical for understanding fiber channel characteristics, developing digital signal processing (DSP) technologies, optimizing optical network configurations, and advancing the optical fiber transmission system towards wideband. Deep learning (DL) has emerged as a powerful tool for waveform modeling, offering high accuracy and low complexity compared to traditional split-step Fourier method (SSFM), due to its strong nonlinear fitting capabilities and efficient parallel computation. However, most DL methods are designed for few-channel and low-rate WDM systems, leaving their scalability to wideband systems uncertain. Moreover, the lack of a standardized accuracy evaluation method and the inconsistent results between waveform errors and transmission performance errors, hinders fair comparisons of various DL schemes. In this paper, we introduce a DSP-assisted accuracy evaluation method integrated with nonlinear DSP, providing a fair benchmark for evaluating the accuracy of DL models. Using this method, we conduct a comprehensive comparison of DL schemes, ranging from simple configurations to more complex wideband setups. The feature decoupled distributed method combining with bidirectional long short-term memory achieves the better performance compared to other DL schemes. Furthermore, in scenarios with more-channel and higher-rate, the performance advantages of FDD-BiLSTM will be further improved. However, as the number of channels and symbol rates increase, the performance of FDD-BiLSTM still gradually deteriorate. We analyze these challenges from three perspectives: the more intricate linear and nonlinear effects, the higher sampling rate required for SSFM. To address these challenges, we discuss potential solutions from two aspects: incorporating more prior physical knowledge and optimizing the structure of DL models.
We demonstrate the first cross-domain cross-layer level-4 autonomous optical network via a multi-AI-agent system. Field trials show ~98% task completion rate across the distributed AI training lifecycle-3.2x higher than single agents using state-of-the-art LLMs.
In this paper, we investigate and assess the performance of intra-channel nonlinearity compensation (IC-NLC) in long-haul coherent optical transmission systems with a symbol rate of 200 GBaud and beyond. We first evaluate the potential gain of ideal IC-NLC in 4 THz systems by estimating the proportion of self-channel interference (SCI) using the split-step Fourier method (SSFM) based simulation with either lumped amplification or distributed amplification. As the symbol rate increases to 300 GBaud, the SCI proportion exceeds 65%. On the other hand, the non-deterministic polarization mode dispersion (PMD) will impact the effectiveness of IC-NLC, especially for ultra-high symbol rate systems. Therefore, we investigate the power spectral density of the residual nonlinear noise after ideal IC-NLC in the presence of PMD. The results indicate that the gain of ideal digital backpropagation (IDBP) decreases by 3.85 dB in 300 GBaud erbium-doped fiber amplifier (EDFA)-amplified links with a PMD parameter of 0.05 ps/km1/2, and 5.09 dB in distributed Raman amplifier (DRA)-amplified links. Finally, we evaluate the potential gains of practical IC-NLC in C-band wavelength-division multiplexing (WDM) systems by employing the low-pass-filter assisted digital backpropagation (LDBP). As the symbol rate increases from 100 GBaud to 300 GBaud, the gain of 20-step-per-span (20-stps) LDBP increases from 0.53 dB to 0.87 dB for EDFA-amplified links, and from 0.89 dB to 1.30 dB for DRA-amplified links. Our quantitative results show that for 200 GBaud and beyond systems, there is a sizable gain to achieve by compensating for intra-channel nonlinearity even with a large non-deterministic PMD.
This paper addresses the challenges of monitoring optical-fiber channels subject to complex, multidimensional impairments-such as dynamic interference across polarization or modal dimensions-where conventional methods suffer from high equipment costs, poor impairment discrimination and limited scalability. We propose an in-service, frequency-domain joint monitoring scheme based on constant-amplitude zero-autocorrelation (CAZAC) sequences. Exploiting their flat spectra and ideal autocorrelation, we model the channel as a multi-input multi-output (MIMO) system and estimate its frequency response to extract both differential group delay (DGD) and dimension-dependent loss (DL) regardless of dimensionality. Experimental validation in polarization-division-multiplexing (PDM) and mode-division-multiplexing (MDM) scenarios demonstrates robust performance: in a 2x2 PDM setup, polarization-dependent loss (PDL) error stays below 0.3 dB and polarization-mode dispersion (PMD) accuracy is 0.3 ps; in a 4x4 MDM system, mode-dependent loss (MDL) and differential mode-group delay (DMGD) errors remain around 0.3 dB and 0.3 ps, respectively. Fully compatible with existing coherent DSP without additional hardware, the scheme enables continuous, cost-effective, real-time monitoring of multidimensional optical channels.
We propose a novel input refinement methodology incorporating sensitivity analysis and memory-aware weighting for jointly refining numerous diverse inputs. Field trials show ~2.5 dB and ~2.3 dB improvements in Q-factor and power estimation, respectively.
We utilize the Feature Decoupling Distributed (FDD) method to enhance the capability of deep learning to fit the Nonlinear Schrodinger Equation (NLSE), significantly reducing the NLSE loss compared to non decoupling model.
We propose and experimentally validate a joint estimation method for chromatic dispersion and time-frequency offset based on the fractional Fourier transform, which reduces computational complexity by more than 50% while keeping estimation accuracy.
We design and demonstrate the first field trial of LLM-powered AI Agent for ADON. Three operation modes of the Agent are proposed for network lifecycle management. The Agent efficiently processes wavelength add/drop and soft/hard failures, and achieves comparable performance to human-designed algorithms for power optimization.
Coherent optics has demonstrated significant potential as a viable solution for achieving 100 Gb/s and higher speeds in single-wavelength passive optical networks (PON). However, upstream burst-mode coherent detection is a major challenge when adopting coherent optics in access networks. To accelerate digital signal processing (DSP) convergence with a minimal preamble length, we propose a novel burst-mode preamble design based on a constant amplitude zero auto-correlation sequence. This design facilitates comprehensive estimation of linear channel effects in the frequency domain, including polarization state rotation, differential group delay, chromatic dispersion, and polarization-dependent loss, providing overall system response information for channel equalization pre-convergence. Additionally, this preamble utilizes the same training unit to jointly achieve three key DSP functions: frame synchronization, frequency offset estimation, and channel estimation. This integration contributes to a significant reduction in the preamble length. The feasibility of the proposed preamble with a length of 272 symbols and corresponding DSP was experimentally verified in a 15 Gbaud coherent system using dual-polarization 16 quadrature amplitude modulation. The experimental results based on this scheme showed a superior performance of the convergence acceleration.
We propose a heuristic-based optimization scheme for reliable optical amplifier reconfiguration process in ADON. In the experiment on a commercial testbed, the scheme prevents a 1.0-dB Q-factor degradation and outperforms 98.5% random solutions.
We propose a n input parameter refinement scheme for the physics-based Raman amplifier model. Experiments over C+L band are conducted. Results show the scheme can lower the physical model's maximum estimation error by 2.13 dB.
We present a high-capacity self-homodyne optical transmission system that enables simultaneously multidimensional vibration sensing based on a weakly-coupled 7-core fiber. To our knowledge, we demonstrate for the first-time detection of fiber vibration direction along with strength, frequency, and location of the vibration source, while transmitting in the meantime single-carrier 16 QAM signal reaching a net date rate of 5.36 Tb/s over 41.4 km of telecom 7-core fiber.
To enable intelligent and self-driving optical networks, high-accuracy physical layer models are required. The dynamic wavelength-dependent gain effects of non-constant-pump erbium-doped fiber amplifiers (EDFAs) remain a crucial problem in terms of modeling, as it determines optical-to-signal noise ratio as well as the magnitude of fiber nonlinearities. Black-box data-driven models have been widely studied, but it requires a large size of data for training and suffers from poor generalizability. In this paper, we derive the gain spectra of EDFAs as a simple univariable linear function, and then based on it we propose a grey-box EDFA gain modeling scheme. Experimental results show that for both automatic gain control (AGC) and automatic power control (APC) EDFAs, our model built with 8 data samples can achieve better performance than the neural network (NN) based model built with 900 data samples, which means the required data size for modeling can be reduced by at least two orders of magnitude. Moreover, in the experiment the proposed model demonstrates superior generalizability to unseen scenarios since it is based on the underlying physics of EDFAs. The results indicate that building a customized digital twin of each EDFA in optical networks become feasible, which is essential especially for next generation multi-band network operations.