Deep neural networks with large model sizes achieve state-of-the-art results for tasks in computer vision and natural language processing. However, such models are too compute- or memory-intensive for resource-constrained edge devices. Prior works on parallel and distributed execution primarily focus on training-rather than inference-using homogeneous accelerators in data centers. We propose PipeEdge, a distributed framework for edge systems that uses pipeline parallelism to both speed up inference and enable running larger, more accurate models that otherwise cannot fit on single edge devices. PipeEdge uses an optimal partition strategy that considers heterogeneity in compute, memory, and network bandwidth. Our empirical evaluation demonstrates that PipeEdge achieves 11.88× and 12.78× speedup using 16 edge devices for the ViT-Huge and BERT-Large models, respectively, with no accuracy loss. Similarly, PipeEdge improves throughput for ViT-Huge (which cannot fit in a single device) by 3.93× over a 4-device baseline using 16 edge devices. Finally, we show up to 4.16× throughput improvement over the state-of-the-art PipeDream when using a heterogeneous set of devices.
Pipeline parallelism has achieved great success in deploying large-scale transformer models in cloud environments, but has received less attention in edge environments. Unlike in cloud scenarios with high-speed and stable network interconnects, dynamic bandwidth in edge systems can degrade distributed pipeline performance. We address this issue with QuantPipe, a communication-efficient distributed edge system that introduces post-training quantization (PTQ) to compress the communicated tensors. QuantPipe uses adaptive PTQ to change bitwidths in response to bandwidth dynamics, maintaining transformer pipeline performance while incurring limited inference accuracy loss. We further improve the accuracy with a directed-search analytical clipping for integer quantization method (DS-ACIQ), which bridges the gap between estimated and real data distributions. Experimental results show that QuantPipe adapts to dynamic bandwidth to maintain pipeline performance while achieving a practical model accuracy using a wide range of quantization bitwidths, e.g., improving accuracy under 2-bit quantization by 15.85\% on ImageNet compared to naive quantization.
Deep learning is commonly used to make personalized recommendations to users for a wide variety of activities. However, deep learning recommendation model (DLRM) training is increasingly dominated by all-to-all and many-to-many communication patterns. While there are a wide variety of algorithms to efficiently overlap communication and computation for many collective operations, these patterns are strictly limited by network bottlenecks. We propose co-designing DLRM model training with the recently proposed Opera network, which is designed to avoid multiple network hops using time-varying source-to-destination circuits. Using measurements from state-of-the-art NVIDIA A100 GPUs, we simulate DLRM model training on networks ranging from 16 to 1024 nodes and demonstrate up to 1.79× improvement using Opera compared with equivalent fat-tree networks. We identify important parameters affecting training time and demonstrate that careful co-design is needed to optimize training latency.