Weiqiang Sun

Niangen Ye, Jiawen Zhu, Baojun Chen et al.

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.

Niangen Ye, Jingya Liu, Weiqiang Sun et al.

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.

Zida Liang, Jiayi Zhu, Weiqiang Sun

Transformer-based architectures achieved high performance in natural language processing and computer vision, yet many studies have shown that they have not demonstrated a clear advantage in time series forecasting and even underperform simple linear baselines in some cases. However, most of these studies have not thoroughly explored the reasons behind the failure of transformers. To better understand time-series transformers(TST), we designed a series of experiments, progressively modifying transformers into MLPs to investigate the impact of the attention mechanism. Surprisingly, transformer blocks often degenerate into simple MLPs in existing time-series transformers. We designed a interpretable dataset to investigate the reasons behind the failure of the attention mechanism and revealed that the attention mechanism is not working in the expected way. We theoretically analyzed the reasons behind this phenomenon, demonstrating that the current embedding methods fail to allow transformers to function in a well-structured latent space, and further analyzed the deeper underlying causes of the failure of embedding.

Zida Liang, Changfa Wu, Dunxian Huang et al.

Recommendation systems are essential tools in modern e-commerce, facilitating personalized user experiences by suggesting relevant products. Recent advancements in generative models have demonstrated potential in enhancing recommendation systems; however, these models often exhibit limitations in optimizing retrieval tasks, primarily due to their reliance on autoregressive generation mechanisms. Conventional approaches introduce sequential dependencies that impede efficient retrieval, as they are inherently unsuitable for generating multiple items without positional constraints within a single request session. To address these limitations, we propose TBGRecall, a framework integrating Next Session Prediction (NSP), designed to enhance generative retrieval models for e-commerce applications. Our framework reformulation involves partitioning input samples into multi-session sequences, where each sequence comprises a session token followed by a set of item tokens, and then further incorporate multiple optimizations tailored to the generative task in retrieval scenarios. In terms of training methodology, our pipeline integrates limited historical data pre-training with stochastic partial incremental training, significantly improving training efficiency and emphasizing the superiority of data recency over sheer data volume. Our extensive experiments, conducted on public benchmarks alongside a large-scale industrial dataset from TaoBao, show TBGRecall outperforms the state-of-the-art recommendation methods, and exhibits a clear scaling law trend. Ultimately, NSP represents a significant advancement in the effectiveness of generative recommendation systems for e-commerce applications.

Xiaoyi Zhou, Liang Huang, Tong Ye et al.

This paper investigates an unmanned aerial vehicle (UAV)-assisted wireless powered mobile-edge computing (MEC) system, where the UAV powers the mobile terminals by wireless power transfer (WPT) and provides computation service for them. We aim to maximize the computation rate of terminals while ensuring fairness among them. Considering the random trajectories of mobile terminals, we propose a soft actor-critic (SAC)-based UAV trajectory planning and resource allocation (SAC-TR) algorithm, which combines off-policy and maximum entropy reinforcement learning to promote the convergence of the algorithm. We design the reward as a heterogeneous function of computation rate, fairness, and reaching of destination. Simulation results show that SAC-TR can quickly adapt to varying network environments and outperform representative benchmarks in a variety of situations.