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.

The inference workload redistribution is a technique for evacuating inference requests from hot edges to idle edges in edge collaborative systems, thereby achieving inference workload balancing for inference on different edges. However, with the continuous development of edge accelerators, the resource utilization of edge accelerators in executing inference requests in series is often low, and when executing multiple inference requests in parallel, it faces uncertain execution delays, different response-time Service Level Objectives (SLOs), and the generality of inference workloads in heterogeneous edge collaborative systems. To address these issues, for the first time in the domain of inference workload redistribution, we propose a Batch-aware Inference workload Redistribution and Parallel execution scheme, called BIRP, to reduce the additional latency caused by waiting for a single inference task during serial execution, thereby improving the overall inference accuracy. BIRP uses the Multi-Armed Bandit (MAB) algorithm to adjust hyperparameters of the Throughput Improvement Ratio (TIR) function online for improving the overall inference accuracy. For nonlinear terms in the problem, BIRP uses a piecewise linear approximation to convert it into a Quadratic Programming (QP) problem, ensuring the effectiveness of BIRP in theory. We prototype BIRP on an edge collaborative system composed of three heterogeneous edges. Based on real inference workload trace, we validate the superiority of our algorithm compared to the state-of-the-art model selection-based inference workload redistribution algorithm, with an overall inference loss reduction of at least 32.9% and the failure rate of SLO has been reduced to 19.8% of alternatives.

Federated learning is emerging as a major learning paradigm, which enables multiple devices to train a model col-laboratively and to keep the privacy of data. However, substantial computation-intensive iterations are performed on devices before the training completion, which incurs heavy consumption of the energy. Along with the stabilization of those model parameters being trained, such on-device training iterations are redundant gradually over time. Thus, we propose to scale the update results obtained from reduced iterations as the substitute for on-device training, based on current model status and device heterogeneity. We thus formulate a time-varying integer program, to minimize cumulative energy consumption over devices, subject to a long-term constraint regarding the model convergence. We then design a polynomial-time online algorithm upon system dynamics, which essentially balances the energy consumption and the model quality being trained. Via rigorous proofs, our approach only incurs sub linear regret, compared with its optimum, and ensures related model convergence. Extensive testbed experiments for real training confirm the superiority of our approach, over multiple alternatives, under various scenarios, decreasing at least 30.2% energy consumption, while preserving the accuracy of the model.