Task scheduling is a critical component in edge computing. Many advanced scheduling strategies select the most suitable execution node based on user-reported resource requirements, thereby enhancing user experience. However, the reliance on user-reported resource demands presents a potential vulnerability, as malicious users could exploit this to disrupt edge servers. In this paper, we propose a hypothetical offloading attack scenario in which an attacker submits tasks with falsified resource demands, directing these tasks to specific nodes. A high volume of malicious tasks leads to congestion at the targeted node, causing delays and potential failures for latency-sensitive tasks. Due to the black-box nature of the offloading strategy and the unknown system environment, we model the task parameter manipulation as a multi-dimensional multi-armed bandit (MAB) process with non-stationary rewards. This approach aims to maximize the attack efficiency within limited attack costs. We introduce UCB algorithm for Multidimensional space with Cost constraints and Non-stationary rewards(UCB-MCN), an extended MAB algorithm designed to optimize the attacker's cumulative reward. In a simulated environment, UCB-MCN is tested against various offloading strategies and defense mechanisms. Evaluation results demonstrate the effectiveness of UCB-MCN and highlight new security concerns within edge computing.

Distributed training has become a cornerstone of large-scale deep learning, yet gradient synchronization remains a critical bottleneck—particularly in high packet loss environments such as wide-area networks (WANs), where frequent retransmissions lead to long-tail latency and degraded training performance. In this paper, we propose a novel approach that leverages spatiotemporal gradient importance modeling to optimize both the transmission protocol and switch queue dropping strategy. Our method dynamically evaluates the importance of each gradient based on its position in the model layers and the current training stage, enabling adaptive packet loss control. In terms of implementation, we integrate a hybrid UDP/TCP transmission protocol: UDP ensures high-throughput gradient delivery, while TCP carries control signals to provide reliable loss feedback. Additionally, our system uses DSCP marking to map gradients to different switch queues, with each queue configured with a distinct WRED (Weighted Random Early Detection) loss threshold that is periodically updated based on the dynamically assessed gradient importance. Extensive experiments demonstrate that our model effectively captures the intrinsic spatiotemporal variations in gradient behavior and significantly reduces long-tail latency while maintaining model accuracy. This provides a robust framework for optimizing gradient synchronization and network resource allocation in distributed deep learning systems.

Advancements in distributed machine learning have led to a demand for engineers with knowledge in AI, hardware, and networking. Research aims to automate deployment based on user intent to improve development efficiency. Large Language Models (LLMs) assist in translating user intent for resource allocation but face challenges in translation accuracy and contextual integrity. The Chain-of-Thought (CoT) improves LLM reasoning by breaking down problems into sequential steps. However, the reasoning steps and dependencies between the model requirements and the computation and network configurations are complex, requiring the selection of an appropriate thought path to construct the CoT framework. To address these issues, we propose a computation and network resource deployment intent translation framework based on CoT reasoning and create a benchmark for user intent in distributed learning. This framework translates user intent into hardware and network configurations, using LLMs to optimize translation accuracy through logical reasoning. Experiments on four LLM models show significant accuracy improvements with CoT compared to traditional prompts. The feasibility of our framework has been validated through its implementation and testing on a real-world testbed.

Large-scale incast traffic generated by AI training tasks poses significant challenges to RDMA networks. Existing congestion control mechanisms, such as Artificial Intelligence ECN (AI ECN) and fixed-ratio PFC, struggle to mitigate frequent PFC pauses caused by ECN's delayed feedback. This paper proposes DSCC, a synergistic algorithm of PFC and ECN. Based on the AI ECN algorithm, DSCC adjusts the PFC threshold according to the incast degree. The adjustment process adheres to the threshold constraints of PFC and ECN. Experiments show that DSCC can significantly improve network performance.

Large Language Models (LLMs) are increasingly being explored across a range of reasoning tasks. However, LLMs sometimes struggle with reasoning tasks under uncertainty that are relatively easy for humans, such as planning actions in stochastic environments. The adoption of LLMs for reasoning is impeded by uncertainty challenges, such as LLM uncertainty and environmental uncertainty. LLM uncertainty arises from the stochastic sampling process inherent to LLMs. Most LLM-based Decision-Making (LDM) approaches address LLM uncertainty through multiple reasoning chains or search trees. However, these approaches overlook environmental uncertainty, which leads to poor performance in environments with stochastic state transitions. Some recent LDM approaches deal with uncertainty by forecasting the probability of unknown variables. However, they are not designed for multi-step reasoning tasks that require interaction with the environment. To address uncertainty in LLM decision-making, we introduce PlanU, an LLM-based planning method that captures uncertainty within Monte Carlo Tree Search (MCTS). PlanU models the return of each node in the MCTS as a quantile distribution, which uses a set of quantiles to represent the return distribution. To balance exploration and exploitation during tree search, PlanU introduces an Upper Confidence Bounds with Curiosity (UCC) score which estimates the uncertainty of MCTS nodes. Through extensive experiments, we demonstrate the effectiveness of PlanU in LLM-based reasoning tasks under uncertainty.

The increasing demand for resource-efficient AI training has positioned Mobile Edge Computing (MEC) as a pivotal component in distributed machine learning tasks. Traditional scheduling methods, however, often fail to account for the inherent characteristics of training tasks, such as periodicity and task correlation, leading to sub-optimal resource utilization. To address these limitations, we propose the Multi-Graph to Graph Scheduler (MG2GS), a novel framework designed to optimize resource allocation and scheduling in MEC environments. MG2GS employs a graph neural network-based feature extractor to capture the spatiotemporal availability of network resources, while a reinforcement learning-based scheduler ensures optimal task scheduling decisions. By incorporating task structure and long-term resource distribution, MG2GS enhances resource utilization by more than 20% compared to existing methods. Simulation results demonstrate its effectiveness in increasing the number of scheduled tasks and improving overall resource efficiency for distributed AI model training.