Rethinking Zero-Loss in AI Data Centers: Towards Bounded-Loss Reliability in Deep Learning Systems
As modern machine learning (ML) workloads grow in size and complexity, communication has emerged as a critical systems bottleneck—not only during training, but across the full lifecycle of ML computation, including inference, pretraining, and fine-tuning: Distributed execution frameworks routinely rely on synchronized communication for exchanging gradients, intermediate activations, token embeddings, routing metadata, and attention context, all of which are vulnerable to tail latency, especially in large-scale environments. Delays in any one of these exchanges can stall iteration progress, degrade throughput, and underutilize expensive compute (e.g., GPU) resources. This dissertation rethinks communication in ML systems through the lens of domain-specific loss tolerance, and proposes two complementary systems that leverage the statistical resilience of ML workloads to partial information, offering a new approach to reducing latency and improving scalability. The first, OptiReduce, is a software-based solution targeting collective communication during distributed data parallel training, the most commonly used distributed ML technique. It introduces a best-effort, time-bounded AllReduce mechanism that tolerates bounded gradient loss to bypass stragglers and reduce iteration stalls. OptiReduce contributes three key techniques: (1) a Transpose AllReduce (TAR) algorithm that establishes direct peerto- peer communication among nodes, avoiding the loss propagation inherent in ring-based collectives where errors accumulate through intermediate aggregations; (2) an Unreliable Bounded Transport (UBT) with adaptive timeouts that caps communication latency regardless of network variability, using dynamic incast control and early timeout strategies to maximize gradient delivery within each time window; and (3) Hadamard Transform encoding that disperses the impact of dropped gradients across entire tensors, converting localized packet loss into uniformly distributed noise that ML models can absorb. Across public cloud environments including CloudLab, OptiReduce achieves up to 70% and 30% faster timeto- accuracy compared to Gloo and NCCL collective communication libraries, respectively, while limiting gradient loss to under 0.1% and maintaining model convergence accuracy even under lossy or congested network conditions. The second, OptiNIC, generalizes these principles to the RDMA transport layer in hardware. To extend beyond AllReduce to all collective patterns, support both training and inference workloads, and meet the 100–400 Gbps line-rate requirements of modern ML clusters, OptiNIC offloads the core ideas of bounded, best-effort communication directly into the NIC. It eliminates retransmissions and in-order delivery from the NIC datapath, enabling a best-effort, out-of-order transport model for RDMA. Unlike traditional transports that signal completion only after full data delivery, OptiNIC introduces bounded completion semantics—where operations complete within application-specified timeouts, triggering forward progress even when data is missing. By removing reorder buffers, retransmission queues, and per-packet sequencing logic, OptiNIC cuts NIC BRAM usage by 2.7× and nearly doubles mean-time-between-failure. Loss recovery shifts to lightweight software mechanisms such as the Hadamard Transform, while standard congestion control (e.g., DCQCN, EQDS, Swift) remains intact. Across public cloud deployments, OptiNIC improves time-to-accuracy by 2× for training, increases inference throughput by 1.6×, and reduces 99th-percentile latency by 3.5×—all while preserving model accuracy. Together, these systems demonstrate that domain-specific collective and transport designs, tailored to ML’s unique statistical properties, can unlock new levels of efficiency, resilience, and scalability in distributed machine learning.
Read more
1
