Large Language Model (LLM) training increasingly relies on Remote Direct Memory Access (RDMA) to enable ultra-efficient networking. However, the unique traffic characteristics—sparse yet bandwidth-intensive—often lead to severe load imbalance under Equal-Cost Multi-Path (ECMP) routing. Packet Spraying (PS) offers a promising solution by distributing traffic across multiple paths, but its impact on congestion dynamics remains insufficiently studied. This paper presents a comprehensive study of PS in Artificial Intelligence (AI) clusters using NS-3 simulations, analyzing its effects on congestion distribution, packet reordering, and flow completion time. Our findings show that congestion patterns vary significantly with workload intensity and oversubscription ratios, and existing congestion control schemes are inadequate for general PS networks, where both routing paths and congestion hotspots frequently change. To address this gap, we propose APSCC, a congestion control algorithm that infers congestion locations from out-of-order packets and aggregates Explicit Congestion Notification (ECN) signals across paths for precise rate adaptation. Compared to state-of-the-art mechanisms, APSCC reduces Job Completion Time (JCT) by up to 30 %. The implementation is publicly available at https://github.com/tangjianback/APSCC.

Tobacco is one of the most important cash crops in China, and accurate acquisition of its planting information is of great significance for tobacco management and yield prediction. In this paper, a tobacco information extraction method based on UAV high-resolution images is proposed. The method uses the RtinaNet target detection model to achieve the identification and localization of tobacco in UAV high-resolution images, and generates tobacco coordinate information. Based on the tobacco coordinates, we designed two algorithms for extracting plant spacing, row spacing and area of tobacco fields, respectively. The experimental results showed that the accuracy of the method reached 98.49%-99.77%, 97.08%-99.98% and 94.48%-99.63% for area, row spacing and plant spacing extracted from the seven experimental tobacco fields, respectively.

Soil plant analysis development (SPAD) represents relative chlorophyll content, which directly affects tabacco health. Accurate monitoring of tobacco canopy SPAD is vital to guide field management. Due to low correlation, few number of features and single model, the existed inversion models have low accuracy and poor robustness. This paper expanded samples from hyperspectral images using PROSAIL, so as to avoid overfitting. In the aspect of feature extraction, the optimized vegetation index is added to increase the correlation between features and SPAD. The inversion model combines K-means and XGBoost to form a mixed model. The results show that the mixed model has better effect on the validation set than other models, R2=0.83, RMSE=3.9.

Image crowd counting is a challenging problem. This study proposes a new deep learning method that estimates crowd counting for the congested scene. The proposed network is composed of two major components: the first ten layers of VGG16 are used as the backbone network, and a dual‐branch (named as Branch_S and Branch_D) network is proposed to be the second part of the network. Branch_S extracts low‐level information (head blob) through a shallow fully convolutional network and Branch_D uses a deep fully convolutional network to extract high‐level context features (faces and body). Features learnt from the two different branches can handle the problem of scale variation due to perspective effects and image size differences. Features of different scales extracted from the two branches are fused to generate predicted density map. On the basis of the fact that an original graph must contain more or equal number of persons than any of its sub‐images, a ranking loss function utilising the constraint relationship inside an image is proposed. Moreover, the ranking loss is combined with Euclidean loss as the final loss function. Our approach is evaluated on three benchmark datasets, and better results are achieved compared with the state‐of‐the‐art works.

Predictable network performance is critical for cloud applications and can be achieved by providing tenants a dedicated virtual data center (VDC) with bandwidth guarantee. Recently, the extended Hose model was applied to the VDC abstraction to characterize the tradeoff between cost and network performance. The acceptability determination problem of a VDC with heterogeneous bandwidth demand was proved to be NP-complete, even in the simple tree topology. In this paper, we investigate the embedding problem for heterogeneous bandwidth VDC in substrate networks of general topology. The embedding problem involves two coupled sub-problems: virtual machine (VM) placement and multipath route assignment. First, we formulate the route assignment problem with linear programming to minimize the maximum link utilization, and provide K-widest path load-balanced routing with controllable splitting paths. Next, we propose a polynomial-time heuristic algorithm, referred to as the perturbation algorithm, for the VM placement. The perturbation algorithm is congestion-aware as it detects the bandwidth bottlenecks in the placement process and then selectively relocates some assigned VMs to eliminate congestion. Simulation results show that our algorithm performs better in comparison with the existing well-known algorithms: first-fit, next-fit, and greedy, and very close to the exponential-time complexity backtracking algorithm in typical data center network architectures. For the tree substrate network, the perturbation algorithm performs better than the allocation-range algorithm. For the homogeneous bandwidth VDC requests, the perturbation algorithm produces a higher success rate than the recently proposed HVC-ACE algorithm. Therefore, it provides a compromised solution between time complexity and network performance.

Network virtualization supports predicted network performance for applications by providing tenants with a virtual data center in multi-tenant data centers. The Hose model was recently extended for the virtual network abstraction and deployed in an Oktopus system, which offers the trade-off between cost and network performance. Embedding algorithms based on Oktopus model were well studied in the single-root tree topology. In this paper, we investigate congestion-aware allocation of virtual data centers in multipath networks. First, we formulate bandwidth constraints with linear programming, and provide a complete solution with exhaustive searching. Next, to reduce time complexity, we propose a perturbation algorithm to VM placement, which detects the bandwidth bottleneck of the current virtual machine placement and then adjusts the assignment to reduce congestion. The perturbation algorithm is compatible with both load-balanced and single-path routing algorithms. We compare the performance of the exhaustive searching algorithm and perturbation algorithm with load-balanced or single-path routing by simulations. The perturbation algorithm with load-balanced routing performs close to the exhaustive searching algorithm while significantly reduces the time complexity. Therefore, it offers a good tradeoff between time complexity and network performance.

In this paper, we investigated designing multicast-capable optical cross-connects (MC-OXCs) on a basis of the 4- stage multicast network. Firstly, we derive the sufficient wide- sense nonblocking (WSNB) and rearrangeable nonblocking (RNB) conditions for the 4-stage multicast network with only two (the second and output) stages being multicast-capable. Both WSNB and RNB 4-stage multicast networks need 0(N3/2) crosspoints. Then MC-OXCs empoying the WSNB and RNB 4-stage multicast network are proposed and proven to be power efficient in reducing the total power loss caused by light splitting.