Graph-based Retrieval-Augmented Generation (GraphRAG) has emerged as a promising paradigm for enhancing LLM reliability by enabling multi-hop reasoning over graph-structured knowledge. However, existing LLMs struggle to efficiently process graph-structured inputs, as traditional attention mechanisms are sequence-based and introduce significant redundancy when serializing graphs into prompt sequences, leading to excessive computation and memory overhead. To address this, we introduce dependency attention, a novel graph-aware attention mechanism that restricts attention computation to token pairs with structural dependencies in the retrieved subgraph. Unlike standard self-attention that computes fully connected interactions, dependency attention prunes irrelevant token pairs and reuses computations along shared relational paths, substantially reducing inference overhead. Building on this idea, we develop DepCache, a KV cache management framework tailored for dependency attention. DepCache enables efficient KV cache reuse through (i) a graph-based KV cache reuse strategy that aligns KV caches across varying prompt contexts, enabling efficient cross-request reuse in GraphRAG, and (ii) a locality-aware replacement policy that leverages spatial and temporal access patterns to improve KV cache hit rate. Evaluations across diverse models and datasets show that DepCache improves LLM inference throughput by 1.5×-5.0× and reduces time-to-first-token latency by up to 3.2×, without compromising generation accuracy.

Graph neural networks (GNNs) have emerged as a promising method for learning from graph data, but large-scale GNN training requires extensive memory and computation resources. To address this, researchers have proposed using multi-GPU processing, which partitions graph data across GPUs for parallel training. However, vertex dependencies in multi-GPU GNN training lead to significant neighbor replications across GPUs, increasing memory consumption. The substantial intermediate data generated during training further exacerbates this issue. Neighbor replication and intermediate data constitute the primary memory consumption in GNN training (i.e., typically accounting for over 80%). In this work, we propose GNN task parallelism for multi-GPU GNN training, which reduces neighbor replication by partitioning training tasks in each layer across different GPUs rather than partitioning the graph structure. This approach only partitions the graph data within individual GPUs, reducing the memory requirements of single tasks while overlapping subgraph computation across different GPUs. Shared neighbor embeddings among different subgraphs can be efficiently reused within a single GPU. Additionally, we employ a task-decoupled GNN training framework, which decouples different training tasks to manage their associated intermediate data independently and release it as early as possible to reduce memory usage. By integrating these techniques, we propose a multi-GPU GNN training system, NeutronTask. Experimental results on a 4×A5000 GPU server show that NeutronTask effectively supports billion-scale full-graph GNN training. For small graphs where the training data fits into the GPUs, NeutronTask achieves 1.27× - 5.47× speedup compared to state-of-the-art GNN systems including NeutronStar and Sancus.

Graph Neural Networks (GNNs) have shown exceptional performance across a wide range of applications. Current frameworks leverage CPU-GPU heterogeneous environments for GNN model training, incorporating mini-batch and sampling techniques to mitigate GPU memory constraints. In such settings, sample-based GNN training can be divided into three phases: sampling, gathering, and training. Existing GNN systems deploy various task orchestration methods to execute each phase on either the CPU or GPU. However, through comprehensive experimentation and analysis, we observe that these task orchestration approaches do not optimally exploit the available heterogeneous resources, hindered by either inefficient CPU processing or GPU resource bottlenecks. In this paper, we propose NeutronOrch, a system for sample-based GNN training that ensures balanced utilization of the CPU and GPU. NeutronOrch decouples the training process by layer and pushes down the training task of the bottom layer to the CPU. This significantly reduces the computational load and memory footprint of GPU training. To avoid inefficient CPU processing, NeutronOrch only offloads the training of frequently accessed vertices to the CPU and lets GPU reuse their embeddings with bounded staleness. Furthermore, NeutronOrch provides a fine-grained pipeline design for the layer-based task orchestrating method. The experimental results show that compared with the state-of-the-art GNN systems, NeutronOrch can achieve up to 11.51× performance speedup.

Graph neural networks (GNNs) have emerged as a promising direction. Training large-scale graphs that relies on distributed computing power poses new challenges. Existing distributed GNN systems leverage data parallelism by partitioning the input graph and distributing it to multiple workers. However, due to the irregular nature of the graph structure, existing distributed approaches suffer from unbalanced workloads and high overhead in managing cross-worker vertex dependencies. In this paper, we leverage tensor parallelism for distributed GNN training. GNN tensor parallelism eliminates cross-worker vertex dependencies by partitioning features instead of graph structures. Different workers are assigned training tasks on different feature slices with the same dimensional size, leading to a complete load balance. We achieve efficient GNN tensor parallelism through two critical functions. Firstly, we employ a generalized decoupled training framework to decouple NN operations from graph aggregation operations, significantly reducing the communication overhead caused by NN operations which must be computed using complete features. Secondly, we employ a memory-efficient task scheduling strategy to support the training of large graphs exceeding single GPU memory, while further improving performance by overlapping communication and computation. By integrating the above techniques, we propose a distributed GNN training system NeutronTP. Our experimental results on a 16-node Aliyun cluster demonstrate that NeutronTP achieves 1.29×-8.72× speedup over state-of-the-art GNN systems including DistDGL, NeutronStar, and Sancus.

Many Graph Neural Network (GNN) training systems have emerged recently to support efficient GNN training. Since GNNs embody complex data dependencies between training samples, the training of GNNs should address distinct challenges different from DNN training in data management, such as data partitioning, batch preparation for mini-batch training, and data transferring between CPUs and GPUs. These factors, which take up a large proportion of training time, make data management in GNN training more significant. This paper reviews GNN training from a data management perspective and provides a comprehensive analysis and evaluation of the representative approaches. We conduct extensive experiments on various benchmark datasets and show many interesting and valuable results. We also provide some practical tips learned from these experiments, which are helpful for designing GNN training systems in the future.

Processing large graphs with memory-limited GPU needs to resolve issues of host-GPU data transfer, which is a key performance bottleneck. Existing GPU-accelerated graph processing frameworks reduce the data transfers by managing the active subgraph transfer at runtime. Some frameworks adopt explicit transfer management approaches based on explicit memory copy with filter or compaction. In contrast, others adopt implicit transfer management approaches based on on-demand access with zero-copy or unified-memory. Having made intensive analysis, we find that as the active vertices evolve, the performance of the two approaches varies in different workloads. Due to heavy redundant data transfers, high CPU compaction overhead, or low bandwidth utilization, adopting a single approach often results in suboptimal performance.In this work, we propose a hybrid transfer management approach to take the merits of both the two approaches at runtime, with an objective to achieve the shortest execution time in each iteration. Based on the hybrid approach, we present HyTGraph, a GPU-accelerated graph processing framework, which is empowered by a set of effective task scheduling optimizations to improve the performance. Our experimental results on real-world and synthesized graphs demonstrate that HyTGraph achieves up to 10.27X speedup over existing GPU-accelerated graph processing systems including Grus, Subway, and EMOGI.