The rapid adoption of Large Language Models (LLMs) in cloud environments has intensified the demand for high-performance AI training and inference, where Remote Direct Memory Access (RDMA) plays a critical role. However, existing RDMA virtualization solutions, such as Single-Root Input/Output Virtualization (SR-IOV), face significant limitations in scalability, performance, and stability. These issues include lengthy container initialization times, hardware resource constraints, and inefficient traffic steering. To address these challenges, we propose Stellar, a new generation RDMA network for cloud AI. Stellar introduces three key innovations: Para-Virtualized Direct Memory Access (PVDMA) for on-demand memory pinning, extended Memory Translation Table (eMTT) for optimized GPU Direct RDMA (GDR) performance, and RDMA Packet Spray for efficient multi-path utilization. Deployed in our large-scale AI clusters, Stellar spins up virtual devices in seconds, reduces container initialization time by 15 times, and improves LLM training speed by up to 14%. Our evaluations demonstrate that Stellar significantly outperforms existing solutions, offering a scalable, stable, and high-performance RDMA network for cloud AI.

In this paper, an optimization framework of partial differential neural network combined with divide-space sampling (PDNN-DSS) is proposed to solve the estimation and control problem of 3D diffusion-advection systems. The PDNN-DSS framework contains a divide-space sampling observer, a model prediction estimator, and a partial differential neural network solver. First, the open-loop state of the diffusion-advection system is obtained by designing the division space observer. The observer provides a deployment scheme for sensors so that mobile robots only need to carry actuators not sensors. Then, the control problem is transformed into an optimization problem by designing a model prediction estimator. Subsequently, a novel partial differential neural network is designed. The neural network is based on partial differential equations containing spatial diffusion features and can solve spatio-temporal optimization problems. Finally, the effectiveness of the proposed method in this paper is demonstrated through simulations from multi-dimensional perspectives of 2D and 3D diffusion-advection systems.

In this article, a state-coupled neural network (SDNN) is proposed to solve the distributed multiple robotic arms (DMRAs) synchronous collaboration problem. The synchronized collaboration of DMRAs is not only in the Cartesian space of the end-effector but also in the corresponding joint velocity space to keep the joint velocity synchronized. First, the constraints for motion generation of leader and follower robots are obtained based on the desired trajectory and communication topology, respectively. Then, the DMRAs collaboration is transformed into quadratic programming based on the minimum velocity norm index. Second, a novel SDNN is designed based on the communication topology of the DMRAs to solve the quadratic programming problem, and the stability of the SDNN is proved by the Lyapunov method. Finally, simulations and experiments demonstrate that SDNN can solve the synchronized collaboration problem of DMRAs with unique advantages.

To avoid the task failure caused by joint breakdown during the collaborative motion planning of dual-redundant robot manipulators, a neural dynamic fault-tolerant (NDFT) scheme is proposed and applied. To do so, a joint fault-tolerant strategy is first designed, and it is formulated as a time-varying equality constraint. Second, combining the robot position and orientation control, joint limit constraint, joint fault-tolerant equality constraint, and considering the repetitive motion optimization criterion, a fault-tolerant framework for the dual-redundant robot manipulators based on quadratic programming (QP) is constructed. Then, a varying-parameter recurrent neural network (VP-RNN) is designed to solve the QP issue. The fault-tolerant framework and the VP-RNN constitute NDFT scheme. With the NDFT scheme, the impact of faulty joints on the whole system can be remedied by healthy joints, thereby the end-effectors of the robot can complete the given end-effector task. Finally, computer simulations and physical experiments are implemented to verify the availability, physical realizability, and accuracy of the proposed NDFT scheme in the collaborative execution of end-effector tasks. Comparative experimental results with conventional repetitive motion planning schemes based on neural networks show higher accuracy and smaller joint angle drift.

Solid State Drives (SSDs) are widely used in data-intensive scenarios due to their high performance and decreasing cost. However, in shared environments, concurrent workloads can interfere with each other, leading to a violation of Quality of Service (QoS). While QoS mechanisms like fairness guarantees and latency constraints have been integrated into SSDs, existing transaction processing frameworks offer limited QoS guarantees and can significantly degrade overall performance in a shared environment. The reason is that the internal components of an SSD, originally designed to exploit parallelism, struggle to coordinate effectively when QoS mechanisms are applied to them. This article proposes a novel QoS -enhanced transaction pro cessing framework, called QoS-pro, which enhances QoS guarantees for concurrent workloads while maintaining high parallelism for SSDs. QoS-pro achieves this by redesigning transaction processing procedures to fully exploit the parallelism of shared SSDs and enhancing QoS-oriented transaction translation and scheduling with parallelism features in mind. In terms of fairness guarantees, QoS-pro outperforms state-of-the-art methods by achieving 96% fairness improvement and 64% maximum latency reduction. QoS-pro also shows almost no loss in throughput when compared with parallelism-oriented methods. Additionally, QoS-pro triggers the fewest Garbage Collection (GC) operations and minimally affects concurrently running workloads during GC operations.

A swarm-exploring neurodynamic network (SENN) based on a two-timescale model is proposed in this study for solving nonconvex nonlinear programming problems. First, by using a convergent-differential neural network (CDNN) as a local quadratic programming (QP) solver and combining it with a two-timescale model design method, a two-timescale convergent-differential (TTCD) model is exploited, and its stability is analyzed and described in detail. Second, swarm exploration neurodynamics are incorporated into the TTCD model to obtain an SENN with global search capabilities. Finally, the feasibility of the proposed SENN is demonstrated via simulation, and the superiority of the SENN is exhibited through a comparison with existing collaborative neurodynamics methods. The advantage of the SENN is that it only needs a single recurrent neural network (RNN) interact, while the compared collaborative neurodynamic approach (CNA) involves multiple RNN runs.

A state estimator based on a time-space sampled-data mechanism is proposed for reaction-diffusion neural networks with uncertain parameters over a rectangular domain$\Omega $. The specific sampling strategy is to establish a coordinate system for the two-dimensional space, divide the two coordinate axis into finite sampling intervals, and then take the midpoints of the respective sampling intervals as the coordinates of the sampling points. The objective is to further reduce the burden of communication by decreasing the number of spatial sampling points while maintaining satisfactory estimation performance. Sufficient conditions for the error system’s stability and the convergence region of quasi-estimation are derived by the Lyapunov function method and the improved Halanay’s inequality. Three numerical examples illustrate the validity and advantage of the proposed method.

In this paper, state observation of coupled reaction-diffusion genetic regulatory networks (GRNs) with time-varying delays is investigated under Dirichlet boundary conditions. First, the above GRNs are constructed to model gene regulatory properties, where the feedback regulation function of the GRNs is assumed to exhibit the Hill form and a novel method to deal with it is introduced. Then a time-space sampled-data state observer is designed for the mentioned networks and new criteria are established by utilizing the Lyapunov stability theory and the inequality techniques of Halanay et al. Finally, the validity of the theoretical results is proved by numerical examples.