Offloading large language models (LLMs) states to host memory during inference promises to reduce operational costs by supporting larger models, longer prompts, and larger batch sizes. However, the design of existing memory offloading mechanisms does not take latency service-level objectives (SLOs) into consideration. As a result, they either lead to frequent SLO violations or underutilize host memory, thereby incurring economic loss and thus defeating the purpose of memory offloading.

Graph neural networks (GNNs) are gaining popularity in diverse application domains and growing in complexity.As a result, it is crucial to achieve high-performance GNN execution.Among various techniques, algebraic transformations, including operator reordering and operator fusion, have been successfully applied to improve the computation and memory access efficiencies of DNN models.However,

Recent foundation models are capable of handling multiple tasks and multiple data modalities with the unified base model structure and several specialized model components. However, efficient training of such multi-task (MT) multi-modal (MM) models poses significant system challenges due to the sophisticated model architecture and the heterogeneous workloads of different tasks and modalities. In this paper, we propose Spindle, a brand new training system tailored for resource-efficient and high-performance training of MT MM models via wavefront scheduling. The key idea of Spindle is to decompose the model execution into waves and address the joint optimization problem sequentially, including both heterogeneity-aware workload parallelization and dependency-driven execution scheduling. We build our system and evaluate it on various MT MM models. Experiments demonstrate the superior performance and efficiency of Spindle, with speedup ratio up to 71% compared to state-of-the-art training systems.

Graph neural networks (GNNs) have proven to be powerful tools for learning from graph-structured data and have achieved great success in many applications. As the sizes of real-world graphs continue to grow, traditional GNN training methods face significant scalability challenges. Recently, disks have gained attention as a cost-effective solution to store large-scale graphs, and several disk-based GNN systems have been proposed to train large-scale graphs on a single machine. However, these systems either overlook the unique data characteristics of GNN workloads when designing cache plans or fail to fully exploit the multilevel hierarchy of storage and computation in system execution, thus resulting in disk I/O bottleneck and resource under-utilization. To address these issues, we present CaliEX, an advanced disk-based GNN system that employs joint optimizations of caching and execution within and across different training stages. CaliEX first designs tailored cache plans and execution policy for both graph topology and features to accelerate neighborhood sampling and feature gathering. Since these two training stages work on different types of data, CaliEX further auto-tunes the cache allocation and pipelines the execution across different stages to improve resource utilization and overall training throughput. Evaluations on multiple GNN models and various large-scale datasets show that CaliEX achieves 3.28 × speedup on average compared to existing disk-based GNN training systems.

Long context fine-tuning of large language models(LLMs) involves training on datasets that are predominantly composed of short sequences and a small proportion of longer sequences. However, existing approaches overlook this long-tail distribution and employ training strategies designed specifically for long sequences. Moreover, these approaches also fail to address the challenges posed by variable sequence lengths during distributed training, such as load imbalance in data parallelism and severe pipeline bubbles in pipeline parallelism. These issues lead to suboptimal training performance and poor GPU resource utilization. To tackle these problems, we propose a chunk-centric training method named ChunkFlow. ChunkFlow reorganizes input sequences into uniformly sized chunks by consolidating short sequences and splitting longer ones. This approach achieves optimal computational efficiency and balance among training inputs. Additionally, ChunkFlow incorporates a state-aware chunk scheduling mechanism to ensure that the peak memory usage during training is primarily determined by the chunk size rather than the maximum sequence length in the dataset. Integrating this scheduling mechanism with existing pipeline scheduling algorithms further enhances the performance of distributed training. Experimental results demonstrate that, compared with Megatron-LM, ChunkFlow can be up to 4.53x faster in the long context fine-tuning of LLMs. Furthermore, we believe that ChunkFlow serves as an effective solution for a broader range of scenarios, such as long context continual pre-training, where datasets contain variable-length sequences.

Tensor parallelism is an essential technique for distributed training of large neural networks. However, automatically determining an optimal tensor parallel strategy is challenging due to the gigantic search space, which grows exponentially with model size and tensor dimension. This prohibits the adoption of auto-parallel systems on larger models.

Long-context supervised fine-tuning (Long-SFT) plays a vital role in enhancing the performance of large language models (LLMs) on long-context tasks. To smoothly adapt LLMs to long-context scenarios, this process typically entails training on mixed datasets containing both long and short sequences. However, this heterogeneous sequence length distribution poses significant challenges for existing training systems, as they fail to simultaneously achieve high training efficiency for both long and short sequences, resulting in sub-optimal end-to-end system performance in Long-SFT. In this paper, we present a novel perspective on data scheduling to address the challenges posed by the heterogeneous data distributions in Long-SFT. We propose Skrull, a dynamic data scheduler specifically designed for efficient long-SFT. Through dynamic data scheduling, Skrull balances the computation requirements of long and short sequences, improving overall training efficiency. Furthermore, we formulate the scheduling process as a joint optimization problem and thoroughly analyze the trade-offs involved. Based on those analysis, Skrull employs a lightweight scheduling algorithm to achieve near-zero cost online scheduling in Long-SFT. Finally, we implement Skrull upon DeepSpeed, a state-of-the-art distributed training system for LLMs. Experimental results demonstrate that Skrull outperforms DeepSpeed by 3.76x on average (up to 7.54x) in real-world long-SFT scenarios.

Online GNN inference has been widely explored by applications such as online recommendation and financial fraud detection systems, where even minor delays can result in significant financial impact. Real-time dynamic graph sampling enables online GNN inference to reflect the latest graph updates in real-world graphs. However, online GNN inference typically demands millisecond-level latency Service Level Objectives (SLOs) as its performance guarantees, which poses great challenges for existing dynamic graph sampling approaches based on graph databases. The issues mainly arise from two aspects: long tail latency due to imbalanced data-dependent sampling and large communication overhead incurred by distributed sampling. To address these issues, we propose Helios, an efficient distributed dynamic graph sampling service to meet the stringent latency SLOs. The key ideas of Helios are 1) pre-sampling the dynamic graph in an event-driven approach, and 2) maintaining a query-aware sample cache to build the complete K-hop sampling results locally for inference requests. Experiments on multiple datasets show that Helios achieves up to 67× higher serving throughput and up to 32× lower P99 query latency compared to baselines.

As deep learning (DL) technologies become ubiquitous, GPU clusters are deployed for inference tasks with consistent service level objectives (SLOs). Efficiently utilizing multiple GPUs is crucial for throughput and cost-effectiveness. This article addresses the challenges posed by dynamic input and NVIDIA MIG in scheduling DL workloads. We present ElasticBatch, a scheduling system that simplifies configuration through bucketization and employs a machine learning-based pipeline to optimize settings. Our experiments demonstrate that ElasticBatch achieves a 50% reduction in GPU instances compared to MIG disablement, increases GPU utilization by 1.4% to 6.5% over an ideal scheduler and significantly reduces profiling time. This research contributes to the discourse on efficient utilization of GPU clusters. ElasticBatch's effectiveness in mitigating challenges posed by dynamic inputs and NVIDIA MIG underscores its potential to optimize GPU cluster performance, providing tangible benefits in terms of reduced instances, increased utilization, and significant time savings in real-world deployment scenarios.

As graph neural networks (GNNs) have achieved great success in many graph learning problems, it is of paramount importance to support their efficient execution. Different graphs and different operators present different patterns during execution. However, there is still a gap in the existing GNN acceleration research to explore adaptive parallelism. We show that existing GNN frameworks rely on handwritten static kernels, which fail to achieve the best performance across different graph operators and input graph structures. In this work, we propose uGrapher, a unified interface that achieves general high performance for different graph operators and datasets. The existing GNN frameworks can easily integrate our design for its simple and unified API. We take a principled approach that decouples a graph operator’s computation and schedule to achieve that. We first build a GNN-specific operator abstraction that incorporates the semantics of graph tensors and graph loops. We explore various schedule strategies based on the abstraction that can balance the well-established trade-off relationship between parallelism, locality, and efficiency. Our evaluation shows that uGrapher can bring up to 29.1× (3.5× on average) performance improvement over the state-of-the-art baselines on two studied NVIDIA GPUs.

Distributed synchronized GPU training is commonly used for deep learning. The resource constraint of using a fixed number of GPUs makes large-scale training jobs suffer from long queuing time for resource allocation, and lowers the cluster utilization. Adapting to resource elasticity can alleviate this but often introduces inconsistent model accuracy, due to lacking of capability to decouple model training procedure from resource allocation. We propose EasyScale, an elastic training system that achieves consistent model accuracy under resource elasticity for both homogeneous and heterogeneous GPUs. EasyScale preserves the data-parallel training behaviors strictly, traces the consistency-relevant factors carefully, utilizes the deep learning characteristics for EasyScaleThread abstraction and fast context-switching. To utilize heterogeneous cluster, EasyScale dynamically assigns workers based on the intra-/inter-job schedulers, minimizing load imbalance and maximizing aggregated job throughput. Deployed in an online serving cluster, EasyScale powers the training jobs to utilize idle GPUs opportunistically, improving overall cluster utilization by 62.1%.

Training data pre-processing pipelines are essential to deep learning (DL). As the performance of model training keeps increasing with both hardware advancements (e.g., faster GPUs) and various software optimizations, the data pre-processing on CPUs is becoming more resource-intensive and a severe bottleneck of the pipeline. This problem is even worse in the cloud, where training jobs exhibit diverse CPU-GPU demands that usually result in mismatches with fixed hardware configurations and resource fragmentation, degrading both training performance and cluster utilization. We introduce GoldMiner, an input data processing service for stateless operations used in pre-processing data for DL model training. GoldMiner decouples data pre-processing from model training into a new role called the data worker. Data workers facilitate scaling of data pre-processing to anywhere in a cluster, effectively pooling the resources across the cluster to satisfy the diverse requirements of training jobs. GoldMiner achieves this decoupling in a fully automatic and elastic manner. The key insight is that data pre-processing is inherently stateless, thus can be executed independently and elastically. This insight guides GoldMiner to automatically extract stateless computation out of a monolithic training program, efficiently disaggregate it across data workers, and elastically scale data workers to tune the resource allocations across jobs to optimize cluster efficiency. We have applied GoldMiner to industrial workloads, and our evaluation shows that GoldMiner can transform unmodified training programs to use data workers, accelerating individual training jobs by up to 12.1x. GoldMiner also improves average job completion time and aggregate GPU utilization by up to 2.5x and 2.1x in a 64-GPU cluster, respectively, by scheduling data workers with elasticity.

With the fast growth of parameter size, it becomes increasingly challenging to deploy large generative models as they typically require large GPU memory consumption and massive computation. Unstructured model pruning has been a common approach to reduce both GPU memory footprint and the overall computation while retaining good model accuracy. However, the existing solutions do not provide an efficient support for handling unstructured sparsity on modern GPUs, especially on the highly-structured tensor core hardware. Therefore, we propose Flash-LLM for enabling low-cost and highly efficient large generative model inference with the sophisticated support of unstructured sparsity on high-performance but highly restrictive tensor cores. Based on our key observation that the main bottleneck of generative model inference is the several skinny matrix multiplications for which tensor cores would be significantly under-utilized due to low computational intensity, we propose a general Load-as-Sparse and Compute-as-Dense methodology for unstructured sparse matrix multiplication (SpMM). The basic insight is to address the significant memory bandwidth bottleneck while tolerating redundant computations that are not critical for end-to-end performance on tensor cores. Based on this, we design an effective software framework for tensor core based unstructured SpMM, leveraging on-chip resources for efficient sparse data extraction and computation/memory-access overlapping. Extensive evaluations demonstrate that (1) at SpMM kernel level, Flash-LLM significantly outperforms the state-of-the-art library, i.e., Sputnik and SparTA by an average of 2.9X and 1.5X, respectively.(2) At end-to-end framework level on OPT-30B/66B/175B models, for tokens per GPU-second , Flash-LLM achieves up to 3.8X and 3.6X improvement over DeepSpeed and FasterTransformer, respectively, with significantly lower inference cost.

The development of personalized recommendation has significantly improved the accuracy of information matching and the revenue of e-commerce platforms. Recently, it has two trends: 1) recommender systems must be trained timely to cope with ever-growing new products and ever-changing user interests from online marketing and social network; 2) state-of-the-art recommendation models introduce deep neural network (DNN) modules to improve prediction accuracy. Traditional CPU-based recommender systems cannot meet these two trends, and GPU-centric training has become a trending approach. However, we observe that GPU devices in training recommender systems are underutilized, and they cannot attain an expected throughput improvement as what it has achieved in Computer Vision (CV) and Neural Language Processing (NLP) areas. This issue can be explained by two characteristics of these recommendation models: First, they contain up to a thousand of input feature fields, introducing fragmentary and memory-intensive operations; Second, the multiple constituent feature interaction submodules introduce substantial small-sized compute kernels. To remove this roadblock to the development of recommender systems, we propose a novel framework named PICASSO to accelerate the training of recommendation models on commodity hardware. Specifically, we conduct a systematic analysis to reveal the bottlenecks encountered in training recommendation models. We leverage the model structure and data distribution to unleash the potential of hardware through our packing, interleaving, and caching optimization. Experiments show that PICASSO increases the hardware utilization by an order of magnitude on the basis of state-of-the-art baselines and brings up to 6× throughput improvement for a variety of industrial recommendation models. Using the same hardware budget in production, PICASSO on average shortens the walltime of daily training tasks by 7 hours, significantly reducing the delay of continuous delivery.

An increasing number of machine learning tasks require dealing with large graph datasets, which capture rich and complex relationship among potentially billions of elements. Graph Neural Network (GNN) becomes an effective way to address the graph learning problem by converting the graph data into a low dimensional space while keeping both the structural and property information to the maximum extent and constructing a neural network for training and referencing. However, it is challenging to provide an efficient graph storage and computation capabilities to facilitate GNN training and enable development of new GNN algorithms. In this paper, we present a comprehensive graph neural network system, namely AliGraph , which consists of distributed graph storage, optimized sampling operators and runtime to efficiently support not only existing popular GNNs but also a series of in-house developed ones for different scenarios. The system is currently deployed at Alibaba to support a variety of business scenarios, including product recommendation and personalized search at Alibaba's E-Commerce platform. By conducting extensive experiments on a real-world dataset with 492.90 million vertices, 6.82 billion edges and rich attributes, AliGraph performs an order of magnitude faster in terms of graph building (5 minutes vs hours reported from the state-of-the-art PowerGraph platform). At training, AliGraph runs 40%-50% faster with the novel caching strategy and demonstrates around 12 times speed up with the improved runtime. In addition, our in-house developed GNN models all showcase their statistically significant superiorities in terms of both effectiveness and efficiency (e.g., 4.12%--17.19% lift by F1 scores).

In past decades, we have witnessed lots of gas leakage diasters all over the world, causing serious casualties, property damage and severe negative social impact. During evacuation after gas leakage incident, the poisonous gas shall be accurately detected. Accordingly, the evacuation routine shall be carefully planned and warned to the evacuating people. Unmanned aerial vehicle (UAV) has been widely regarded as a promising tool to support crowd evacuation. In this paper, aiming at providing an efficient UAV control system for crowd evacuation, we propose a crowd simulation based UAV control system to direct crowd evacuation from the polluted area. The architecture mainly consists of a UAV fleet management module, UAV trajectory planning module, and a crowd simulation module. The architecture forms a closed control loop, emphasizing the inter-operation between the real scenario and the simulation scenario. A case study on UAV gas leakage detection scheduling is given. The results show that the proposed architecture can provide efficient way to help pedestrian in the environment to evacuate and avoid to be infected by the toxic gas.