图形神经网络（GNNS）将深度神经网络（DNN）的成功扩展到非欧几里德图数据，实现了各种任务的接地性能，例如节点分类和图形属性预测。尽管如此，现有系统效率低，培训数十亿节点和GPU的节点和边缘训练大图。主要瓶颈是准备GPU数据的过程 - 子图采样和特征检索。本文提出了一个分布式GNN培训系统的BGL，旨在解决一些关键思想的瓶颈。首先，我们提出了一种动态缓存引擎，以最小化特征检索流量。通过协同设计缓存政策和抽样顺序，我们发现低开销和高缓存命中率的精美斑点。其次，我们改善了曲线图分区算法，以减少子图采样期间的交叉分区通信。最后，仔细资源隔离减少了不同数据预处理阶段之间的争用。关于各种GNN模型和大图数据集的广泛实验表明，BGL平均明显优于现有的GNN训练系统20.68倍。

Graph neural networks (GNNs) have been demonstrated to be a powerful algorithmic model in broad application fields for their effectiveness in learning over graphs. To scale GNN training up for large-scale and ever-growing graphs, the most promising solution is distributed training which distributes the workload of training across multiple computing nodes. However, the workflows, computational patterns, communication patterns, and optimization techniques of distributed GNN training remain preliminarily understood. In this paper, we provide a comprehensive survey of distributed GNN training by investigating various optimization techniques used in distributed GNN training. First, distributed GNN training is classified into several categories according to their workflows. In addition, their computational patterns and communication patterns, as well as the optimization techniques proposed by recent work are introduced. Second, the software frameworks and hardware platforms of distributed GNN training are also introduced for a deeper understanding. Third, distributed GNN training is compared with distributed training of deep neural networks, emphasizing the uniqueness of distributed GNN training. Finally, interesting issues and opportunities in this field are discussed.

图形神经网络（GNN）已被证明是分析非欧国人图数据的强大工具。但是，缺乏有效的分布图学习（GL）系统极大地阻碍了GNN的应用，尤其是当图形大且GNN相对深时。本文中，我们提出了GraphTheta，这是一种以顶点为中心的图形编程模型实现的新颖分布式和可扩展的GL系统。 GraphTheta是第一个基于分布式图处理的GL系统，其神经网络运算符以用户定义的功能实现。该系统支持多种培训策略，并在分布式（虚拟）机器上启用高度可扩展的大图学习。为了促进图形卷积实现，GraphTheta提出了一个名为NN-Tgar的新的GL抽象，以弥合图形处理和图形深度学习之间的差距。提出了分布式图引擎，以通过混合平行执行进行随机梯度下降优化。此外，除了全球批次和迷你批次外，我们还为新的集群批次培训策略提供了支持。我们使用许多网络大小的数据集评估GraphTheta，范围从小，适度到大规模。实验结果表明，GraphTheta可以很好地扩展到1,024名工人，用于培训内部开发的GNN，该工业尺度的Aripay数据集为14亿个节点和41亿个属性边缘，并带有CPU虚拟机（Dockers）群的小群。 （5 $ \ sim $ 12GB）。此外，GraphTheta比最先进的GNN实现获得了可比或更好的预测结果，证明其学习GNN和现有框架的能力，并且可以超过多达$ 2.02 \ tims $ $ 2.02 \ times $，具有更好的可扩展性。据我们所知，这项工作介绍了文献中最大的边缘属性GNN学习任务。

最近，Graph神经网络（GNNS）已成为聚光灯作为强大的工具，可以有效地在图形结构化数据上执行各种推理任务。随着现实图表的大小继续扩展，GNN训练系统面临可扩展性挑战。分布式培训是一种流行的方法，可以通过扩展CPU节点来应对这一挑战。但是，对基于磁盘的GNN培训的关注不多，该培训可以通过利用NVME SSD等高性能存储设备来以更具成本效益的方式扩展单节点系统。我们观察到，主内存和磁盘之间的数据移动是基于SSD的训练系统中的主要瓶颈，并且常规的GNN训练管道是不错的选择，而无需考虑此开销。因此，我们提出了Ginex，这是第一个基于SSD的GNN训练系统，可以在单台计算机上处​​理数十亿个图形数据集。受到编译器优化的检查员执行模型的启发，Ginex通过分开样品和收集阶段来重组GNN训练管道。这种分离使Ginex能够实现一种可证明的最佳替换算法，即被称为Belady的算法，用于存储器中的Caching特征向量，该算法是I/O访问的主要部分。根据我们对40亿尺度图数据集的评估，Ginex平均比SSD扩展的Pytorch几何得出了2.11倍的训练吞吐量（最大最高2.67倍）。

图形神经网络（GNN）的输入图的大小不断增加，突显了使用多GPU平台的需求。但是，由于计算不平衡和效率较低的通信，现有的多GPU GNN解决方案遭受了劣质性能。为此，我们提出了MGG，这是一种新型的系统设计，可以通过以GPU为中心的软件管道在多GPU平台上加速GNN。 MGG探讨了通过细粒度计算通信管道中隐藏GNN工作负载中远程内存访问延迟的潜力。具体而言，MGG引入了管​​道感知工作负载管理策略和混合数据布局设计，以促进通信局限性重叠。 MGG实现以优化的管道为中心的内核。它包括工作负载交织和基于经经的映射，以进行有效的GPU内核操作管道和专门的内存设计以及优化，以更好地数据访问性能。此外，MGG还结合了轻巧的分析建模和优化启发式方法，以动态提高运行时不同设置的GNN执行性能。全面的实验表明，MGG在各种GNN设置上的最先进的多GPU系统要比最先进的多GPU系统：平均比具有统一虚拟内存设计的多GPU系统快3.65倍，平均比DGCL框架快7.38倍。

图形神经网络（GNNS）在学习从图形结构数据中展示了成功，其中包含欺诈检测，推荐和知识图形推理。然而，培训GNN有效地具有挑战性，因为：1）GPU存储器容量有限，对于大型数据集可能不足，而2）基于图形的数据结构导致不规则的数据访问模式。在这项工作中，我们提供了一种统计分析的方法，并确定了GNN培训前更频繁地访问的数据。我们的数据分层方法不仅利用输入图的结构，而且还从实际GNN训练过程中获得了洞察力，以实现更高的预测结果。通过我们的数据分层方法，我们还提供了一种新的数据放置和访问策略，以进一步最大限度地减少CPU-GPU通信开销。我们还考虑了多GPU GNN培训，我们也展示了我们在多GPU系统中的策略的有效性。评估结果表明，我们的工作将CPU-GPU流量降低了87-95％，并通过数亿节点和数十亿边缘的图表提高了现有解决方案的GNN训练速度。

最近，图形卷积网络（GCNS）已成为用于分析非欧几里德图数据的最先进的算法。然而，实现有效的GCN训练，特别是在大图中挑战。原因是许多折叠的原因：1）GCN训练引发了大量的内存占用。大图中的全批量培训甚至需要数百到数千千兆字节的内存，以缓冲中间数据进行反向传播。 2）GCN培训涉及内存密集型数据减少和计算密集型功能/渐变更新操作。这种异构性质挑战当前的CPU / GPU平台。 3）图形的不规则性和复杂的训练数据流共同增加了提高GCN培训系统效率的难度。本文提出了一种混合架构来解决这些挑战的混合架构。具体地，GCNEAR采用基于DIMM的存储系统，提供易于级别的存储器容量。为了匹配异构性质，我们将GCN培训操作分类为内存密集型减少和计算密集型更新操作。然后，我们卸载将操作减少到DIMM NMES，充分利用高聚合的本地带宽。我们采用具有足够计算能力的CAE来处理更新操作。我们进一步提出了几种优化策略来处理GCN任务的不规则，提高GCNEAR的表现。我们还提出了一种多GCNEAR系统来评估GCNEAR的可扩展性。

Graph Convolutional Networks (GCNs) are extensively utilized for deep learning on graphs. The large data sizes of graphs and their vertex features make scalable training algorithms and distributed memory systems necessary. Since the convolution operation on graphs induces irregular memory access patterns, designing a memory- and communication-efficient parallel algorithm for GCN training poses unique challenges. We propose a highly parallel training algorithm that scales to large processor counts. In our solution, the large adjacency and vertex-feature matrices are partitioned among processors. We exploit the vertex-partitioning of the graph to use non-blocking point-to-point communication operations between processors for better scalability. To further minimize the parallelization overheads, we introduce a sparse matrix partitioning scheme based on a hypergraph partitioning model for full-batch training. We also propose a novel stochastic hypergraph model to encode the expected communication volume in mini-batch training. We show the merits of the hypergraph model, previously unexplored for GCN training, over the standard graph partitioning model which does not accurately encode the communication costs. Experiments performed on real-world graph datasets demonstrate that the proposed algorithms achieve considerable speedups over alternative solutions. The optimizations achieved on communication costs become even more pronounced at high scalability with many processors. The performance benefits are preserved in deeper GCNs having more layers as well as on billion-scale graphs.

While many systems have been developed to train Graph Neural Networks (GNNs), efficient model inference and evaluation remain to be addressed. For instance, using the widely adopted node-wise approach, model evaluation can account for up to 94% of the time in the end-to-end training process due to neighbor explosion, which means that a node accesses its multi-hop neighbors. On the other hand, layer-wise inference avoids the neighbor explosion problem by conducting inference layer by layer such that the nodes only need their one-hop neighbors in each layer. However, implementing layer-wise inference requires substantial engineering efforts because users need to manually decompose a GNN model into layers for computation and split workload into batches to fit into device memory. In this paper, we develop Deep Graph Inference (DGI) -- a system for easy and efficient GNN model inference, which automatically translates the training code of a GNN model for layer-wise execution. DGI is general for various GNN models and different kinds of inference requests, and supports out-of-core execution on large graphs that cannot fit in CPU memory. Experimental results show that DGI consistently outperforms layer-wise inference across different datasets and hardware settings, and the speedup can be over 1,000x.

基础模型正在成为主要的深度学习技术。由于模型参数和训练数据集的大规模，预处理基础模型始终耗时。除了计算密集型外，培训过程还非常密集和沟通密集。这些功能使得需要应用3D并行性，该平行性整合数据并行性，管道模型并行性和张量模型并行性，以实现高训练效率。为了实现这一目标，开发了一些自定义软件框架，例如Megatron-LM和DeepSpeed。但是，当前的3D平行框架仍然符合两个问题：i）它们对模型开发人员不透明，这些开发人员需要手动修改模型以并行化培训。 ii）它们对计算，GPU存储器和网络带宽的利用不足。我们提出了Merak，这是一个自动化的3D并行性深度学习培训框架，并具有高度资源利用。 Merak会自动使用自动模型分区仪部署，该分区仪在模型的代理表示上使用图形sharding算法。 Merak还提出了非侵入性的API，用于通过最小的代码修改来扩展基础模型培训。此外，我们在Merak设计了高性能的3D平行运行时引擎。它使用多种技术来利用可用的培训资源，包括移动的关键路径管道时间表，该计划带来了更高的计算利用率，阶段感知的重新计算，可利用空闲工作者的记忆以及子额定张量的模型并行性，这些模型并联与通信和计算重叠。 64 GPU的实验显示，Merak可以加快在最新的3D平行性框架上，具有1.5、2.5、8.3和20亿的模型框架，最高可达1.42x，1.39x，1.43x和1.61 x分别。

最近提出了基于子图的图表学习（SGRL）来应对规范图神经网络（GNNS）遇到的一些基本挑战，并在许多重要的数据科学应用（例如链接，关系和主题预测）中证明了优势。但是，当前的SGRL方法遇到了可伸缩性问题，因为它们需要为每个培训或测试查询提取子图。扩大规范GNN的最新解决方案可能不适用于SGRL。在这里，我们通过共同设计学习算法及其系统支持，为可扩展的SGRL提出了一种新颖的框架Surel。 Surel采用基于步行的子图表分解，并将步行重新形成子图，从而大大降低了子图提取的冗余并支持并行计算。具有数百万个节点和边缘的六个同质，异质和高阶图的实验证明了Surel的有效性和可扩展性。特别是，与SGRL基线相比，Surel可以实现10 $ \ times $ Quad-Up，具有可比甚至更好的预测性能；与规范GNN相比，Surel可实现50％的预测准确性。

Training Graph Neural Networks, on graphs containing billions of vertices and edges, at scale using minibatch sampling poses a key challenge: strong-scaling graphs and training examples results in lower compute and higher communication volume and potential performance loss. DistGNN-MB employs a novel Historical Embedding Cache combined with compute-communication overlap to address this challenge. On a 32-node (64-socket) cluster of $3^{rd}$ generation Intel Xeon Scalable Processors with 36 cores per socket, DistGNN-MB trains 3-layer GraphSAGE and GAT models on OGBN-Papers100M to convergence with epoch times of 2 seconds and 4.9 seconds, respectively, on 32 compute nodes. At this scale, DistGNN-MB trains GraphSAGE 5.2x faster than the widely-used DistDGL. DistGNN-MB trains GraphSAGE and GAT 10x and 17.2x faster, respectively, as compute nodes scale from 2 to 32.

图表卷积网络（GCNS）已成为最先进的图形学习模型。但是，它可以令人难以置于大图数据集的推断GCNS，这会将其应用于大型实际图表并阻碍更深层更复杂的GCN图形的探讨。这是因为真实世界图可能非常大而稀疏。此外，GCN的节点度倾向于遵循幂律分布，因此具有高度不规则的邻接矩阵，导致数据处理和移动中的禁止低效率，从而显着地限制了可实现的GCN加速效率。为此，本文提出了一种GCN算法和加速器协同设计框架被称为GCOD，其在很大程度上可以缓解上述GCN不规则性并提高GCNS推理效率。具体地，在算法级别上，GCOD集成了分割和征服GCN训练策略，该训练策略将图形偏离在本地邻域中的密集或稀疏，而不会影响模型精度，从而导致（主要）的图形邻接矩阵仅仅是两个级别的工作量并享受大部分增强的规律性，从而轻松加速。在硬件水平上，我们进一步开发了一个具有分离发动机的专用双子加速器，以处理每个上述密集和稀疏工作负载，进一步提高整体利用率和加速效率。广泛的实验和消融研究验证了我们的GCOD始终如一地减少了与CPU，GPU和现有技术GCN加速器相比的15286倍，294倍，7.8倍和2.5倍的加速，包括HYGCN和AWB -GCN分别在保持甚至提高任务准确性的同时。

许多现实世界图包含时域信息。时间图神经网络在生成的动态节点嵌入中捕获时间信息以及结构和上下文信息。研究人员表明，这些嵌入在许多不同的任务中实现了最先进的表现。在这项工作中，我们提出了TGL，这是一个用于大规模脱机时间图神经网络训练的统一框架，用户可以使用简单的配置文件组成各种时间图神经网络。 TGL包括五个主要组件，一个临时采样器，一个邮箱，节点内存模块，存储器更新程序和消息传递引擎。我们设计了临时CSR数据结构和平行采样器，以有效地对颞邻邻居进行制作微型批次。我们提出了一种新颖的随机块调度技术，该技术可以减轻大批量训练时过时的节点存储器的问题。为了解决仅在小规模数据集上评估当前TGNN的局限性，我们介绍了两个具有0.2亿和13亿个时间边缘的大型现实世界数据集。我们在四个具有单个GPU的小规模数据集上评估了TGL的性能，以及两个具有多个GPU的大数据集，用于链接预测和节点分类任务。我们将TGL与五种方法的开源代码进行了比较，并表明TGL平均达到13倍的速度可实现相似或更高的精度。与基准相比，我们的时间平行采样器在多核CPU上平均达到173倍加速。在4-GPU机器上，TGL可以在1-10小时内训练一个超过10亿个时间边缘的时期。据我们所知，这是第一项提出了一个关于多个GPU的大规模时间图神经网络培训的一般框架的工作。

近年来，在平衡（超级）图分配算法的设计和评估中取得了重大进展。我们调查了过去十年的实用算法的趋势，用于平衡（超级）图形分区以及未来的研究方向。我们的工作是对先前有关该主题的调查的更新。特别是，该调查还通过涵盖了超图形分区和流算法来扩展先前的调查，并额外关注并行算法。

嵌入式模型是高维数据的有效学习范例。但是，嵌入模型的一个开放问题是它们的表示（潜在因子）通常会导致大参数空间。我们观察到，现有的分布式训练框架面临嵌入模型的可伸缩性问题，因为从服务器的共享嵌入参数更新和检索共享嵌入参数通常占主导地位培训周期。在本文中，我们提出了一种新的系统框架，可显着提高巨大嵌入模型培训的可扩展性。我们拥抱嵌入的嵌入式作为绩效机会的倾斜流行分布，并利用它来解决具有嵌入缓存的通信瓶颈。为确保缓存跨越一致性，我们将新的一致性模型纳入HET设计，该模型提供了在每嵌入的基础上提供细粒度的一致性保证。与以前的工作相比，只允许读取操作的僵化，HET也利用了写入操作的血液性。六种代表性任务的评估表明，在最先进的基线上，HET达到高达88％的嵌入通信减少和高达20.68倍的性能加速。

Dynamic Graph Neural Networks (DGNNs) have been broadly applied in various real-life applications, such as link prediction and pandemic forecast, to capture both static structural information and temporal characteristics from dynamic graphs. Combining both time-dependent and -independent components, DGNNs manifest substantial parallel computation and data reuse potentials, but suffer from severe memory access inefficiency and data transfer overhead under the canonical one-graph-at-a-time training pattern. To tackle the challenges, we propose PiPAD, a $\underline{\textbf{Pi}}pelined$ and $\underline{\textbf{PA}}rallel$ $\underline{\textbf{D}}GNN$ training framework for the end-to-end performance optimization on GPUs. From both the algorithm and runtime level, PiPAD holistically reconstructs the overall training paradigm from the data organization to computation manner. Capable of processing multiple graph snapshots in parallel, PiPAD eliminates the unnecessary data transmission and alleviates memory access inefficiency to improve the overall performance. Our evaluation across various datasets shows PiPAD achieves $1.22\times$-$9.57\times$ speedup over the state-of-the-art DGNN frameworks on three representative models.

TensorFlow is a machine learning system that operates at large scale and in heterogeneous environments. Tensor-Flow uses dataflow graphs to represent computation, shared state, and the operations that mutate that state. It maps the nodes of a dataflow graph across many machines in a cluster, and within a machine across multiple computational devices, including multicore CPUs, generalpurpose GPUs, and custom-designed ASICs known as Tensor Processing Units (TPUs). This architecture gives flexibility to the application developer: whereas in previous "parameter server" designs the management of shared state is built into the system, TensorFlow enables developers to experiment with novel optimizations and training algorithms. TensorFlow supports a variety of applications, with a focus on training and inference on deep neural networks. Several Google services use TensorFlow in production, we have released it as an open-source project, and it has become widely used for machine learning research. In this paper, we describe the TensorFlow dataflow model and demonstrate the compelling performance that Tensor-Flow achieves for several real-world applications.

Graph neural networks (GNNs) have received great attention due to their success in various graph-related learning tasks. Several GNN frameworks have then been developed for fast and easy implementation of GNN models. Despite their popularity, they are not well documented, and their implementations and system performance have not been well understood. In particular, unlike the traditional GNNs that are trained based on the entire graph in a full-batch manner, recent GNNs have been developed with different graph sampling techniques for mini-batch training of GNNs on large graphs. While they improve the scalability, their training times still depend on the implementations in the frameworks as sampling and its associated operations can introduce non-negligible overhead and computational cost. In addition, it is unknown how much the frameworks are 'eco-friendly' from a green computing perspective. In this paper, we provide an in-depth study of two mainstream GNN frameworks along with three state-of-the-art GNNs to analyze their performance in terms of runtime and power/energy consumption. We conduct extensive benchmark experiments at several different levels and present detailed analysis results and observations, which could be helpful for further improvement and optimization.

Deep learning based recommendation models (DLRM) are widely used in several business critical applications. Training such recommendation models efficiently is challenging primarily because they consist of billions of embedding-based parameters which are often stored remotely leading to significant overheads from embedding access. By profiling existing DLRM training, we observe that only 8.5% of the iteration time is spent in forward/backward pass while the remaining time is spent on embedding and model synchronization. Our key insight in this paper is that access to embeddings have a specific structure and pattern which can be used to accelerate training. We observe that embedding accesses are heavily skewed, with almost 1% of embeddings represent more than 92% of total accesses. Further, we observe that during training we can lookahead at future batches to determine exactly which embeddings will be needed at what iteration in the future. Based on these insight, we propose Bagpipe, a system for training deep recommendation models that uses caching and prefetching to overlap remote embedding accesses with the computation. We designed an Oracle Cacher, a new system component which uses our lookahead algorithm to generate optimal cache update decisions and provide strong consistency guarantees. Our experiments using three datasets and two models shows that our approach provides a speed up of up to 6.2x compared to state of the art baselines, while providing the same convergence and reproducibility guarantees as synchronous training.