BERT-like models have been widely adopted in text mining and web search due to their high accuracy. However, large BERT-like models suffer from inefficient online inference on GPUs for two main reasons. First, their high accuracy relies on large model depth, which linearly increases sequential computation on GPUs. Second, stochastic and dynamic online workloads lead to extra costs due to batching and padding. To address the problem, we present Student Parallelism for efficient GPU inference of BERT-like models under real-world online workloads. At its core, Student Parallelism adopts stacking distillation and boosting ensemble, distilling the original deep model into a group of shallow but virtually stacked student models running in parallel. This enables Student Parallelism to achieve a low model depth (e.g., two layers), and thus low inference latency while maintaining accuracy. In addition, we design adaptive student pruning to adjust the number of students according to the dynamic online workloads. For example, during workload bursts, it can temporarily decrease the number of students with minimal accuracy loss to improve system throughput. Extensive experiments on real-world datasets and workloads show that Student Parallelism achieves up to 4.1× lower latency while maintaining accuracy and up to 22.27× higher throughput during workload bursts.

To optimize large Transformer model training, both efficient parallel computing and advanced data management are indispensable. However, current methods often assume a stable and uniform training workload, neglecting data-induced imbalances-arising from both sampling and packing processes-which can impede training performance. Specifically, data sampling imbalance arises from uneven sequence length distribution of the training data, while data packing imbalance stems from the discrepancy between the linear memory complexity and quadratic time complexity of the attention mechanism. To address these imbalance issues, we develop Hydraulis, which jointly optimizes the parallel strategies and data assignment. For one thing, we introduce large model training with dynamic heterogeneous parallel strategies in response to the sequence length variations within and across training iterations. For another, we devise a two-stage data assignment approach, which strikes a good balance in terms of the training workloads both within and across model replicas. Empirical results demonstrate that Hydraulis outperforms existing systems by 1.32-2.66×. Our source code is available: https://github.com/PKU-DAIR/Hetu.

Nowadays, Large Language Models (LLMs) have been trained using extended context lengths to foster more creative applications. However, long context training poses great challenges considering the constraint of GPU memory. It not only leads to substantial activation memory consumption during training, but also incurs considerable memory fragmentation. To facilitate long context training, existing frameworks have adopted strategies such as recomputation and various forms of parallelisms. Nevertheless, these techniques rely on redundant computation or extensive communication, resulting in low Model FLOPS Utilization (MFU). In this paper, we propose MEMO, a novel LLM training framework designed for fine-grained activation memory management. Given the quadratic scaling of computation and linear scaling of memory with sequence lengths when using FlashAttention, we offload memory-consuming activations to CPU memory after each layer's forward pass and fetch them during the backward pass. To maximize the swapping of activations without hindering computation, and to avoid exhausting limited CPU memory, we implement a token-wise activation recomputation and swapping mechanism. Furthermore, we tackle the memory fragmentation issue by employing a bi-level Mixed Integer Programming (MIP) approach, optimizing memory reuse across transformer layers. Empirical results demonstrate that MEMO achieves an average of 1.97x and 1.80x MFU compared to Megatron-LM and DeepSpeed, respectively. This improvement is attributed to MEMO's ability to minimize memory fragmentation, reduce recomputation and intensive communication, and circumvent the delays associated with the memory reorganization process due to fragmentation. By leveraging fine-grained activation memory management, MEMO facilitates efficient training of 7B LLM with 1 million sequence length on just 8 A800 GPUs, achieving an MFU of 52.30%.

The growing memory demands of embedding tables in Deep Learning Recommendation Models (DLRMs) pose great challenges for model training and deployment. Existing embedding compression solutions cannot simultaneously achieve memory efficiency, low latency, and adaptability to dynamic data distribution. This article presents CAFE+, a Compact, Adaptive, and Fast Embedding compression framework that meets the above requirements. The design philosophy of CAFE+ is to dynamically allocate more memory to important features and less to unimportant ones. We assign unique embedding to important feature and allow multiple unimportant features sharing one embedding. We propose a fast and lightweight feature monitor, to real-time capture feature importance and report important features. We theoretically analyze the accuracy of our feature monitor and prove the superiority of CAFE+ from the aspect of model convergence. Extensive experiments show CAFE+ outperforms existing embedding compression methods, yielding \(3.94\%\) and \(3.94\%\) superior testing AUC on Criteo Kaggle dataset and CriteoTB dataset at a compression ratio of \(10{,}000\times\) . Building on our conference version [ 114 ], this journal version introduces several novel designs (implicit importance attenuation, adaptive threshold adjustment, and ColdSifter) that enable CAFE+ to more effectively adapt to long-term online learning and achieve better model quality. All codes are available at GitHub [ 112 ].

Sampling-based Graph Neural Networks (GNNs) have become the de facto standard for handling various graph learning tasks on large-scale graphs. As the graph size grows larger and even exceeds the standard host memory size of a single machine, out-of-core sampling-based GNN training has gained attention from the community. For out-of-core sampling-based GNN training, the performance bottleneck is the data preparation process that includes sampling neighbor lists and gathering node features from external storage. Based on this observation, existing out-of-core GNN training frameworks try to accomplish larger percentages of data requests without inquiring the external storage by designing better in-memory caches. However, the enormous overall requested data volume is unchanged under this approach. In this paper, we present a new perspective on reducing the overall requested data volume. Through a quantitative analysis, we find that Neighborhood Redundancy and Temporal Redundancy exist in out-of-core sampling-based GNN training. To reduce these two kinds of data redundancies, we propose OUTRE, an OUT-of-core de-REdundancy GNN training framework. OUTRE incorporates two new designs, partition-based batch construction and historical embedding cache , to reduce the corresponding data redundancies. Moreover, we propose automatic cache space management to automatically organize available memory for different caches. Evaluation results on four public large-scale graph datasets show that OUTRE achieves 1.52× to 3.51× speedup against the SOTA framework.

We propose to design vertex encoding for determinations of no-result edge queries that should not be executed. Edge query is one of the core operations in mainstream graph databases, which is to retrieve the corresponding edges connecting two given vertices. Real-world graphs may be too large to be stored in memory and frequently accessing edge data on disk usually incurs much overhead. Average degree of real-world graph tends to be much less than the vertex number, and edges may not exist in most pairs of vertices. Efficiently avoiding no-result edge query executions will certainly improve performance of graph database. In this paper, we propose a new and important problem for determining no-result edge queries: vertex encoding for edge nonexistence determination (VEND, for short). We build a low dimensional vertex encoding for all vertices, and we can efficiently determine most vertex pairs that are connected by no edges just with their corresponding codes. With VEND, we can utilize in-memory efficient operations to filter no-result disk accesses for edge query. We also design maintenance algorithms for the proposed solution when data updates happen. Extensive experiments on many real-world datasets confirm the ability of our solution on determining a quite high proportion of non-edge vertex pairs, as well as the acceleration for edge queries.

Recent years have witnessed the unprecedented achievements of large-scale pre-trained models, especially Transformer models. Many products and services in Tencent Inc., such as WeChat, QQ, and Tencent Advertisement, have been opted in to gain the power of pre-trained models. In this work, we present Angel-PTM, a productive deep learning system designed for pre-training and fine-tuning Transformer models. Angel-PTM can train extremely large-scale models with hierarchical memory efficiently. The key designs of Angel-PTM are a fine-grained memory management via the Page abstraction and a unified scheduling method that coordinates computations, data movements, and communications. Furthermore, Angel-PTM supports extreme model scaling with SSD storage and implements a lock-free updating mechanism to address the SSD I/O bottlenecks. Experimental results demonstrate that Angel-PTM outperforms existing systems by up to 114.8% in terms of maximum model scale as well as up to 88.9% in terms of training throughput. Additionally, experiments on GPT3-175B and T5-MoE-1.2T models utilizing hundreds of GPUs verify our strong scalability.

Graph neural networks (GNNs) have achieved great success in many graph-based applications. However, the enormous size and high sparsity level of graphs hinder their applications under industrial scenarios. Although some scalable GNNs are proposed for large-scale graphs, they adopt a fixed K-hop neighborhood for each node, thus facing the over-smoothing issue when adopting large propagation depths for nodes within sparse regions. To tackle the above issue, we propose a new GNN architecture --- Graph Attention Multi-Layer Perceptron (GAMLP), which can capture the underlying correlations between different scales of graph knowledge. We have deployed GAMLP in Tencent with the Angel platform, and we further evaluate GAMLP on both real-world datasets and large-scale industrial datasets. Extensive experiments on these 14 graph datasets demonstrate that GAMLP achieves state-of-the-art performance while enjoying high scalability and efficiency. Specifically, it outperforms GAT by 1.3% regarding predictive accuracy on our large-scale Tencent Video dataset while achieving up to 50x training speedup. Besides, it ranks top-1 on both the leaderboards of the largest homogeneous and heterogeneous graph (i.e., ogbn-papers100M and ogbn-mag) of Open Graph Benchmark.

Due to the rising concerns on privacy protection, how to build machine learning (ML) models over different data sources with security guarantees is gaining more popularity. Vertical federated learning (VFL) describes such a case where ML models are built upon the private data of different participated parties that own disjoint features for the same set of instances, which fits many real-world collaborative tasks. Nevertheless, we find that existing solutions for VFL either support limited kinds of input features or suffer from potential data leakage during the federated execution. To this end, this paper aims to investigate both the functionality and security of ML modes in the VFL scenario.

Graph neural networks (GNNs) have achieved state-of-the-art performance in various graph-based tasks. However, as mainstream GNNs are designed based on the neural message passing mechanism, they do not scale well to data size and message passing steps. Although there has been an emerging interest in the design of scalable GNNs, current researches focus on specific GNN design, rather than the general design space, limiting the discovery of potential scalable GNN models. This paper proposes PaSca, a new paradigm and system that offers a principled approach to systemically construct and explore the design space for scalable GNNs, rather than studying individual designs. Through deconstructing the message passing mechanism, PaSca presents a novel Scalable Graph Neural Architecture Paradigm (SGAP), together with a general architecture design space consisting of 150k different designs. Following the paradigm, we implement an auto-search engine that can automatically search well-performing and scalable GNN architectures to balance the trade-off between multiple criteria (e.g., accuracy and efficiency) via multi-objective optimization. Empirical studies on ten benchmark datasets demonstrate that the representative instances (i.e., PaSca-V1, V2, and V3) discovered by our system achieve consistent performance among competitive baselines. Concretely, PaSca-V3 outperforms the state-of-the-art GNN method JK-Net by 0.4% in terms of predictive accuracy on our large industry dataset while achieving up to 28.3 × training speedups.

Nowadays, it is a rapidly rising demand yet challenging issue to run large-scale applications on shared infrastructures such as data centers and clouds with low execution latency and high resource utilization. This paper reports our experience with Swift, a system capable of efficiently running real-time and interactive data processing jobs at cloud scale. Taking directed acyclic graph DAG as the job model, Swift achieves the design goal by three new mechanisms: 1 fine-grained scheduling that can efficiently partition a job into graphlets i.e., sub-graphs based on new shuffle heuristics and that does scheduling in the unit of graphlet, thus avoiding resource fragmentation and waste, 2 adaptive memory-based in-network shuffling that reduces IO overhead and data transfer time by doing shuffle in memory and allowing jobs to select the most efficient way to fulfill shuffling, and 3 lightweight fault tolerance and recovery that only prolong the whole job execution time slightly with the help of timely failure detection and fine-grained failure recovery. Experimental results show that Swift can achieve an average speedup of 2.11× on TPC-H, and 14.18× on Terasort when compared with Spark. Swift has been deployed in production, supporting as many as 140,000 executors and processing millions of jobs per day. Experiments with production traces show that Swift outperforms JetScope and Bubble Execution by 2.44× and 1.23× respectively.

With the ever-evolving concerns on privacy protection, vertical federated learning (FL), where participants own non-overlapping features for the same set of instances, is becoming a heated topic since it enables multiple enterprises to strengthen the machine learning models collaboratively with privacy guarantees. Nevertheless, to achieve privacy preservation, vertical FL algorithms involve complicated training routines and time-consuming cryptography operations, leading to slow training speed.

Embedding models have been an effective learning paradigm for high-dimensional data. However, one open issue of embedding models is that their representations (latent factors) often result in large parameter space. We observe that existing distributed training frameworks face a scalability issue of embedding models since updating and retrieving the shared embedding parameters from servers usually dominates the training cycle. In this paper, we propose HET, a new system framework that significantly improves the scalability of huge embedding model training. We embrace skewed popularity distributions of embeddings as a performance opportunity and leverage it to address the communication bottleneck with an embedding cache. To ensure consistency across the caches, we incorporate a new consistency model into HET design, which provides fine-grained consistency guarantees on a per-embedding basis. Compared to previous work that only allows staleness for read operations, HET also utilizes staleness for write operations. Evaluations on six representative tasks show that HET achieves up to 88% embedding communication reductions and up to 20.68×performance speedup over the state-of-the-art baselines.

Scalability and fault-tolerance are two fundamental challenges for all distributed computing at Internet scale. Despite many recent advances from both academia and industry, these two problems are still far from settled. In this paper, we present Fuxi, a resource management and job scheduling system that is capable of handling the kind of workload at Alibaba where hundreds of terabytes of data are generated and analyzed everyday to help optimize the company's business operations and user experiences. We employ several novel techniques to enable Fuxi to perform efficient scheduling of hundreds of thousands of concurrent tasks over large clusters with thousands of nodes: 1) an incremental resource management protocol that supports multi-dimensional resource allocation and data locality; 2) user-transparent failure recovery where failures of any Fuxi components will not impact the execution of user jobs; and 3) an effective detection mechanism and a multi-level blacklisting scheme that prevents them from affecting job execution. Our evaluation results demonstrate that 95% and 91% scheduled CPU/memory utilization can be fulfilled under synthetic workloads, and Fuxi is capable of achieving 2.36T-B/minute throughput in GraySort. Additionally, the same Fuxi job only experiences approximately 16% slowdown under a 5% fault-injection rate. The slowdown only grows to 20% when we double the fault-injection rate to 10%. Fuxi has been deployed in our production environment since 2009, and it now manages hundreds of thousands of server nodes.