Zhen Zhang

Daiyaan Arfeen, Zhen Zhang, Xinwei Fu et al.

Training Deep Neural Networks (DNNs) with billions of parameters generally involves pipeline-parallel (PP) execution. Unfortunately, PP model training can use GPUs inefficiently, especially at large scale, due to idle GPU time caused by pipeline bubbles, which are often 15-30% and can exceed 60% of the training job's GPU allocation. To improve the GPU utilization of PP model training, this paper describes PipeFill, which fills pipeline bubbles with execution of other pending jobs. By leveraging bubble GPU time, PipeFill reduces the GPU utilization sacrifice associated with scaling-up of large-model training. To context-switch between fill jobs and the main training job with minimal overhead to the main job, and maximize fill job efficiency, PipeFill carefully fits fill job work to measured bubble durations and GPU memory availability, introduces explicit pipeline-bubble instructions, and orchestrates placement and execution of fill jobs in pipeline bubbles. Experiments show that PipeFill can increase overall utilization by up to 63% for GPUs used in large-scale LLM training, with <2% slowdown of the training job, and 5-15% even for low-scale LLM training. For large-scale LLM training on 8K GPUs, the 63% increase translates to up to 2.6K additional GPUs worth of work completed.

Feng Liang, Zhen Zhang, Haifeng Lu et al.

With the rapid growth in the volume of data sets, models, and devices in the domain of deep learning, there is increasing attention on large-scale distributed deep learning. In contrast to traditional distributed deep learning, the large-scale scenario poses new challenges that include fault tolerance, scalability of algorithms and infrastructures, and heterogeneity in data sets, models, and resources. Due to intensive synchronization of models and sharing of data across GPUs and computing nodes during distributed training and inference processes, communication efficiency becomes the bottleneck for achieving high performance at a large scale. This article surveys the literature over the period of 2018-2023 on algorithms and technologies aimed at achieving efficient communication in large-scale distributed deep learning at various levels, including algorithms, frameworks, and infrastructures. Specifically, we first introduce efficient algorithms for model synchronization and communication data compression in the context of large-scale distributed training. Next, we introduce efficient strategies related to resource allocation and task scheduling for use in distributed training and inference. After that, we present the latest technologies pertaining to modern communication infrastructures used in distributed deep learning with a focus on examining the impact of the communication overhead in a large-scale and heterogeneous setting. Finally, we conduct a case study on the distributed training of large language models at a large scale to illustrate how to apply these technologies in real cases. This article aims to offer researchers a comprehensive understanding of the current landscape of large-scale distributed deep learning and to reveal promising future research directions toward communication-efficient solutions in this scope.