The rapid advancement of large machine learning (ML) models has driven universities worldwide to invest heavily in GPU clusters. Effectively sharing these resources among multiple users is essential for maximizing both utilization and accessibility. However, managing shared GPU clusters presents significant challenges, ranging from system configuration to fair resource allocation among users. This paper introduces SING, a full-stack solution tailored to simplify shared GPU cluster management. Aimed at addressing the pressing need for efficient resource sharing with limited staffing, SING enhances operational efficiency by reducing maintenance costs and optimizing resource utilization. We provide a comprehensive overview of its four extensible architectural layers, explore the features of each layer, and share insights from real-world deployment, including usage patterns and incident management strategies. As part of our commitment to advancing shared ML cluster management, we open-source SING's resources to support the development and operation of similar systems.

Due to the high operational complexity and limited deployment scale of lossless RDMA networks, the community has been exploring efficient RDMA communication over lossy fabrics. State-of-the-art (SOTA) lossy RDMA solutions implement a simplified selective repeat mechanism in RDMA NICs (RNICs) to enhance loss recovery efficiency. However, these solutions still face performance challenges, such as unavoidable ECMP hash collisions and excessive retransmission timeouts (RTOs). In this paper, we revisit RDMA reliability with the goals of being independent of PFC, compatible with packet-level load balancing, free from RTO, and friendly to hardware offloading. To this end, we propose DCP, a transport architecture that co-designs both the switch and RNICs, fully meeting the design goals. At its core, DCP-Switch introduces a simple yet effective lossless control plane, which is leveraged by DCP-RNIC to enhance reliability support for high-speed lossy fabrics, primarily including header-only-based retransmission and bitmap-free packet tracking. We prototype DCP-Switch using P4 switch and DCP-RNIC using FPGA. Extensive experiments demonstrate that DCP achieves 1.6× and 2.1× performance improvements, compared to SOTA lossless and lossy RDMA solutions, respectively.

Mixture-of-Expert (MoE) models outperform conventional models by selectively activating different subnets, named experts, on a per-token basis. This gated computation generates dynamic communications that cannot be determined beforehand, challenging the existing GPU interconnects that remain static during distributed training. In this paper, we advocate for a first-of-its-kind system, called MixNet, that unlocks topology reconfiguration during distributed MoE training. Towards this vision, we first perform a production measurement study and show that the MoE dynamic communication pattern has strong locality, alleviating the need for global reconfiguration. Based on this, we design and implement a regionally reconfigurable high-bandwidth domain that augments existing electrical interconnects using optical circuit switching (OCS), achieving scalability while maintaining rapid adaptability. We build a fully functional MixNet prototype with commodity hardware and a customized collective communication runtime. Our prototype trains state-of-the-art MoE models with in-training topology reconfiguration across 32 A100 GPUs. Large-scale packet-level simulations show that MixNet achieves performance comparable to a non-blocking fat-tree fabric while boosting the networking cost efficiency (e.g., performance per dollar) of four representative MoE models by 1.2×–1.5× and 1.9×–2.3× at 100 Gbps and 400 Gbps link bandwidths, respectively.

Training large language models (LLMs) generates diverse coflows within a cluster, requiring optimized scheduling to enhance communication-computation overlap and minimize training time. Existing schedulers inadequately handle contention both across and within coflows, resulting in suboptimal performance.

Secret Sharing (SS) is widely adopted in secure Multi-Party Computation (MPC) with its simplicity and computational efficiency. However, SS-based MPC protocol introduces significant communication overhead due to interactive operations on secret sharings over the network. For instance, training a neural network model with SS-based MPC may incur tens of thousands of communication rounds among parties, making it extremely hard for real-world deployment.To reduce the communication overhead of SS, prior works statically convert interactive operations to equivalent non-interactive operations with extra computation cost. However, we show that such static conversion misses chances for optimization, and further present SOLAR, an SS-based MPC framework that aims to reduce the communication overhead through dynamic communication-computation conversion. At its heart, SOLAR converts interactive operations that involve communication among parties to equivalent non-interactive operations within each party with extra computations and introduces a speculative strategy to perform opportunistic conversion when CPU is idle for network transmission. We have implemented and evaluated SOLAR on several popular MPC applications, and achieved 1.6-8.1 times speedup in multi-thread setting compared to the basic SS and 1.2-8.6 times speedup over static conversion.

With the growing complexity of deep learning applications, users have started to delegate their data and models to the cloud. Among these applications, online learning services, which involve both training and inference procedures, are widely deployed. To ensure privacy guarantee on the public cloud, researchers have proposed a plethora of privacy-preserving deep learning algorithms with different techniques, ranging from obfuscation mechanisms to cryptographic tools. However, none of them is applicable to online learning services. They either focus only on inference or training procedure while ignoring the other, or require non-colluding or trusted third parties. In this paper, we present Sphinx, an efficient and privacy-preserving online deep learning system without any trusted third parties. Sphinx strikes a balance between model performance, computational efficiency, and privacy preservation with systematical optimizations on both private inference and training protocols. At its core, Sphinx synthesizes homomorphic encryption and differential privacy reciprocally to maintain the model by keeping most of its parameters as plaintexts, enabling fast training and inference protocol designs. Meanwhile, by refining the homomorphic operation behaviors, Sphinx avoids most of the heavyweight homomorphic operations and minimizes the communication cost. As a result, Sphinx is able to reduce the training time significantly while achieving real-time inference without exposing user privacy. In our experiments, we find that compared to the pure homomorphic encryption solution, Sphinx is $35 \times$ faster for training and 4 orders of magnitude faster for inference, providing real-time inference response (0.05 seconds for MNIST and 0.08 seconds for CIFAR-10). Our experiments also demonstrate that Sphinx achieves promising model accuracy under a tight privacy budget (96% accuracy under $\epsilon=2, \delta=10^{-5}$ for MNIST) without a trusted data aggregator, and is more robust against practical reconstruction attacks.

Vertical Federated Learning (VFL) enables multiple parties to collaboratively train a machine learning model over vertically distributed datasets without data privacy leakage. However, there is a limitation of the current VFL solutions: current VFL models fail to conduct inference on non-overlapping samples during inference. This limitation seriously damages the VFL model’s availability because, in practice, overlapping samples may only take up a small portion of the whole data at each party which means a large part of inference tasks will fail. In this article, we propose a novel VFL framework which enables federated inference on non-overlapping data. Our framework regards the distributed features as privileged information which is available in the training period but disappears during inference. We distill the knowledge of such privileged features and transfer them to the parties’ local model which only processes local features. Furthermore, we adopt Oblivious Transfer (OT) to preserve data ID privacy during training and inference. Empirically, we evaluate the model on the real-world dataset collected from Criteo and Taobao. Besides, we also provide a security analysis of the proposed framework.