Distributed AI training often suffers from network faults. Network faults, especially at the last hop between a switch and a host, result in loss of connectivity, resulting in training job stalls and eventual failure. This is typically managed through a fail-stop mechanism, followed by a restart, incurring significant inefficiencies. We present ReCCL, the first network fault-tolerant collective communication library (CCL) that allows training progress to be preserved by seamlessly failing over to alternate paths when a network fault occurs. During failover, ReCCL keeps communication states synchronized while using dynamic channel load balancing and intra-host GPU routing to improve communication performance. Our evaluations demonstrate that ReCCL can perform failover seamlessly with minimal performance losses. Additionally, our simulations also demonstrate that failover can be effectively used to achieve significant savings in GPU hours for large-scale distributed AI training workloads.

The development of large language models (LLMs) poses new challenges in data center network topology design. To assist in exploring topology design, we propose ATOP, an Automated Topology Optimization Pipeline, which models network topology as a set of hyperparameters, enabling the discovery of potential topologies. With various optimization algorithms and customizable optimization objectives, ATOP achieves automated topology optimization on a scale of tens of thousands of GPUs. We apply ATOP on network topologies for 256, 1024, 4096, and 16384 GPUs, optimizing performance under LLMs training traffic patterns, collective communication performance, fault tolerance, and network cost. We also evaluate ATOP in different scenarios: building, optimizing, and expanding a data center. From ATOP's results, we discover a new topology — ZCube, which reaches the highest cost-effectiveness across various GPU scales. Simulation results show that ZCube, compared to the previous state-of-the-art topologies, including Rail-optimized Fat-tree (ROFT), Rail-only, and HPN, improves end-to-end LLM training speed by 3% to 7% and reduces network hardware costs by 26% to 46%. We also construct ZCube on a real-world testbed. Results show that ZCube reduces hardware costs by 25% compared to Rail-Optimized Topology while maintaining the same all-reduce and all-to-all performance.

Mixture-of-Experts (MoE) showcases tremendous potential to scale large language models (LLMs) with enhanced performance and reduced computational complexity. However, its sparsely activated architecture shifts feed-forward networks (FFNs) from being compute-intensive to memory-intensive during inference, leading to substantially lower GPU utilization and increased operational costs.

Reliability is essential for ensuring efficiency in LLM training. However, many real-world reliability issues remain difficult to resolve, resulting in wasted resources and degraded model performance. Unfortunately, today's collective communication libraries operate as black boxes, hiding critical information needed for effective root cause analysis.

Deep neural networks (DNNs) have gained tremendous attractions as compelling solutions for applications such as image classification, object detection, speech recognition, and so forth. Its great success comes with excessive trainings to make sure the model accuracy is good enough for those applications. Nowadays, it becomes challenging to train a DNN model because of 1) the model size and data size keep increasing, which usually needs more iterations to train; 2) DNN algorithms evolve rapidly, which requires the training phase to be short for a quick deployment. To address those challenges, distributed training platforms have been proposed to leverage massive server nodes for training, with the hope of significant training time reduction. Therefore, scalability is a critical performance metric to evaluate a distributed training platform. Nevertheless, our analysis reveals that traditional server clusters have poor scalability for training due to the traffic congestions within the server and beyond. The intra-server traffic on the I/O fabric can result in severe congestions and skewed quality of service as high performance devices are competing with each other. Moreover, the traffic congestions on the Ethernet for inter-server communication could also incur significant performance degradation. In this work, we devise a novel distributed training platform, EFLOPS, that adopts an algorithm and system co-design methodology to achieve good scalability. A new server architecture is proposed to alleviate the intra-server congestions. Moreover, a new network topology, BiGraph, is proposed to divide the network into two separate parts, so that there is always a direct connection between any nodes from different parts. Finally, accompany with BiGraph, a topology-aware allreduce algorithm is proposed to eliminate the traffic congestion on the direct connection. The experimental results show that eliminating the congestions on network interface can gain up to 11.3xcommunication speedup. The proposed algorithm and topology can provide further improvement up to 6.08x. The overall performance of ResNet-50 training achieves near-linear scalability, and is competitive to the top-rankings of MLPerf results.

Over the past one and half years, we have been using RDMA over commodity Ethernet (RoCEv2) to support some of Microsoft's highly-reliable, latency-sensitive services. This paper describes the challenges we encountered during the process and the solutions we devised to address them. In order to scale RoCEv2 beyond VLAN, we have designed a DSCP-based priority flow-control (PFC) mechanism to ensure large-scale deployment. We have addressed the safety challenges brought by PFC-induced deadlock (yes, it happened!), RDMA transport livelock, and the NIC PFC pause frame storm problem. We have also built the monitoring and management systems to make sure RDMA works as expected. Our experiences show that the safety and scalability issues of running RoCEv2 at scale can all be addressed, and RDMA can replace TCP for intra data center communications and achieve low latency, low CPU overhead, and high throughput.