PAI灵骏智算服务

产品介绍选择PAI灵骏，您可以轻松构建具有以下优势的智能集群：算力即服务。提供高性能、高弹性异构算力服务，支持万张GPU规模的资源弹性，单集群网络容量4Pbps，时延低至2微秒。高资源效率。资源利用率提升3倍，并行计算效率90%以上。融合算力池。支持AI+HPC场景算力的统一分配和融合调度，无缝连接。算力管理监控。为异构算力深度定制IT运维管理平台，实现异构算力到池化资源到使用效率的全流程监控管理。为什么选择PAI灵骏选择PAI灵骏，您可以轻松构建具有以下优势的智能集群：算力即服务。提供高性能、高弹性异构算力服务，支持万张GPU规模的资源弹性，单集群网络容量4Pbps，时延低至2微秒。高资源效率。资源利用率提升3倍，并行计算效率90%以上。融合算力池。支持AI+HPC场景算力的统一分配和融合调度，无缝连接。算力管理监控。为异构算力深度定制IT运维管理平台，实现异构算力到池化资源到使用效率的全流程监控管理。产品优势加速AI创新。全链路性能提速，计算密集型项目迭代效率可提升2倍以上。最大化ROI。高效的池化异构算力调度技术，确保每一份算力投入都能得到充分利用，资源利用率可提升3倍。无惧规模挑战。轻松应对大模型和大规模工程仿真的算力需求，让创新不受算力限制。可视又可控。简单的管理异构算力的分配，并持续的监控和优化。应用场景PAI灵骏主要面向图形图像识别、自然语言处理、搜索广告推荐、通用大模型等大规模分布式的AI研发场景，适用于自动驾驶、金融风控、药物研发、科学智能、元宇宙、互联网和ISV等行业。大规模分布式训练。超大规模GPU算力系统。全对等网络架构，全资源池化，可以搭配PAI（机器学习平台）使用，支持多种训练框架（Pytorch、TensorFlow、Caffe、Keras、Xgboost、Mxnet等），可以满足多种规模的AI训练和推理业务。AI基础设施。平滑扩容。满足不同规模GPU算力需求，平滑扩容，性能线性拓展。智能数据加速。针对AI训练场景提供数据智能加速，主动预热训练所需数据，提升训练效率。更高资源利用率。支持异构资源细粒度管控，提升资源周转效率。自动驾驶。丰富的部署和调度策略。多种GPU资源调度策略，保证训练任务高效执行。文件存储CPFS（Cloud Paralleled File System）搭配RDMA网络架构，保证训练数据供给和计算IO；并可使用OSS分级存储降低归档数据存储成本。同时支持训练和仿真场景。融合算力智能供应，同时支持训练仿真两种场景，从协同模式上提升迭代效率，降低数据迁移成本。科学智能。拓展提升创新上限。基于数据中心超大规模RDMA“高速网”和通信流控技术，实现端到端微秒级通信时延，超大规模线性拓展可打造万卡级并行算力。融合生态，拓展创新边界。支持HPC和AI任务融合调度，为科研和AI提供统一协同的底座支撑，促进技术生态融合。云上科研，普惠算力。支持云原生和容器化的AI和HPC应用生态，资源深度共享，普惠的智能算力触手可得。功能特性高速RDMA网络架构。阿里巴巴2016年开始投入专项研究RDMA（Remote Direct Memory Access），目前已建成大规模数据中心内的“高速网”，通过大规模RDMA网络部署实践，阿里云自主研发了基于端网协同的RDMA高性能网络协议和HPCC拥塞控制算法，并通过智能网卡实现了协议硬件卸载，降低了端到端网络延时，提升了网络IO吞吐能力，并有效规避和弱化了网络故障、网络黑洞等传统网络异常给上层应用带来的性能损失。高性能集合通信库ACCL。PAI灵骏支持高性能集合通信库ACCL（Alibaba Collective Communication Library），结合硬件（例如：网络交换机），对万卡规模的AI集群提供无拥塞、高性能的集群通讯能力。阿里云通过通信库ACCL实现了GPU和网卡的智能匹配、节点内外物理拓扑自动识别及拓扑感知的无拥塞通信算法，彻底消除网络拥塞，提升网络通信效率，提高分布式训练系统的扩展性。在万卡规模下，可达80%以上的线性集群能力。在百卡规模下，有效（计算）性能可达95%以上，可满足80%以上的业务场景需求。高性能数据主动加载加速软件KSpeed。PAI灵骏基于高性能网络RDMA和高性能通信ACCL，研发高性能数据主动加载加速软件KSpeed，进行智能数据IO优化。计算存储分离架构广泛存在于AI、HPC、大数据业务场景中，但大量训练数据的加载容易形成效率瓶颈。阿里云通过高性能数据主动加载加速软件KSpeed，实现数据IO数量级性能提升。GPU容器虚拟化方案eGPU。针对AI作业规模庞大、GPU硬件资源昂贵、集群GPU利用率低等业务场景实际遇到的问题，PAI灵骏支持GPU虚拟化技术eGPU，可有效提升AI集群的GPU利用率，具体如下：支持显存、算力双维度自由切分。支持多个规格。支持动态创建、销毁。支持热升级。支持用户态技术，保证更高可靠性。

Product Introduction
By choosing PAI Lingjun, you can easily build intelligent clusters with the following advantages:
1. Computing Power as a Service (CPaaS). Provides high-performance, highly elastic heterogeneous computing power services, supporting resource elasticity at the scale of 10,000 GPUs, with a single-cluster network capacity of 4 Pbps and a latency as low as 2 microseconds.
2. High Resource Efficiency. The resource utilization rate is increased by 3 times, and the parallel computing efficiency exceeds 90%.
3. Integrated Computing Power Pool. Supports unified allocation and integrated scheduling of computing power for AI + HPC scenarios, with seamless connectivity.
4. Computing Power Management and Monitoring. Deeply customized IT operation and maintenance management platform for heterogeneous computing power, realizing end-to-end monitoring and management from heterogeneous computing power to pooled resources to usage efficiency.

Why Choose PAI Lingjun By choosing PAI Lingjun, you can easily build intelligent clusters with the following advantages:
1. Computing Power as a Service (CPaaS). Provides high-performance, highly elastic heterogeneous computing power services, supporting resource elasticity at the scale of 10,000 GPUs, with a single-cluster network capacity of 4 Pbps and a latency as low as 2 microseconds.
2. High Resource Efficiency. The resource utilization rate is increased by 3 times, and the parallel computing efficiency exceeds 90%.
3. Integrated Computing Power Pool. Supports unified allocation and integrated scheduling of computing power for AI + HPC scenarios, with seamless connectivity.
4. Computing Power Management and Monitoring. Deeply customized IT operation and maintenance management platform for heterogeneous computing power, realizing end-to-end monitoring and management from heterogeneous computing power to pooled resources to usage efficiency.

Product Advantages
1. Accelerate AI Innovation. Full-stack performance acceleration, with the iteration efficiency of compute-intensive projects increased by more than 2 times.
2. Maximize ROI. Efficient pooled heterogeneous computing power scheduling technology ensures that every computing power investment is fully utilized, and the resource utilization rate can be increased by 3 times.
3. Meet Scale Challenges with Confidence. Easily meet the computing power requirements of large models and large-scale engineering simulations, allowing innovation to be unrestricted by computing power.
4. Visible and Controllable. Simply manage the allocation of heterogeneous computing power, and conduct continuous monitoring and optimization.

Application Scenarios
PAI Lingjun is mainly oriented to large-scale distributed AI R&D scenarios such as graphic and image recognition, natural language processing (NLP), search advertising and recommendation, general-purpose large models, etc., and is applicable to industries such as autonomous driving, financial risk control, drug discovery, scientific intelligence, metaverse, Internet, and Independent Software Vendors (ISV).
1. Large-scale Distributed Training. Ultra-large-scale GPU computing power system. Adopts full peer-to-peer network architecture and full resource pooling, can be used with PAI (Machine Learning Platform), supports multiple training frameworks (Pytorch, TensorFlow, Caffe, Keras, Xgboost, Mxnet, etc.), and can meet AI training and inference services of various scales.
2. AI Infrastructure. Smooth scaling. Meet GPU computing power requirements of different scales, with smooth scaling and linear performance expansion.
3. Intelligent Data Acceleration. Provide intelligent data acceleration for AI training scenarios, actively preheat the data required for training to improve training efficiency.
4. Higher Resource Utilization. Supports fine-grained management of heterogeneous resources to improve resource turnover efficiency.
5. Autonomous Driving. Rich deployment and scheduling strategies. Multiple GPU resource scheduling strategies to ensure efficient execution of training tasks. Match CPFS (Cloud Paralleled File System) with RDMA network architecture to ensure training data supply and computing IO; use OSS tiered storage to reduce archived data storage costs. Support both training and simulation scenarios. Integrated intelligent computing power supply supports both training and simulation scenarios, improving iteration efficiency and reducing data migration costs from the collaboration mode.
6. Scientific Intelligence. Expand the upper limit of innovation. Based on the ultra-large-scale RDMA "high-speed network" and communication flow control technology in the data center, realize end-to-end microsecond-level communication latency, and ultra-large-scale linear expansion can create 10,000-card parallel computing power. Integrated ecosystem to expand the boundaries of innovation. Supports integrated scheduling of HPC and AI tasks, providing a unified collaborative base for scientific research and AI, and promoting the integration of the technical ecosystem. Cloud-based research and inclusive computing power. Supports cloud-native and containerized AI and HPC application ecosystems, with in-depth resource sharing, and inclusive intelligent computing power is readily available.

Functional Features
1. High-speed RDMA Network Architecture. Alibaba started special research on RDMA (Remote Direct Memory Access) in 2016, and has now built an intra-data-center "high-speed network". Through large-scale RDMA network deployment practices, Alibaba Cloud independently developed end-to-network collaboration-based high-performance RDMA network protocols and HPCC congestion control algorithms, and realized protocol hardware offloading through smart NICs, reducing end-to-end network latency, improving network IO throughput, and effectively avoiding and weakening the performance losses caused by traditional network anomalies such as network faults and network black holes to upper-layer applications.
2. High-performance Collective Communication Library ACCL. PAI Lingjun supports the high-performance collective communication library ACCL (Alibaba Collective Communication Library). Combined with hardware (such as network switches), it provides congestion-free, high-performance cluster communication capabilities for AI clusters at the scale of 10,000 cards. Alibaba Cloud realizes intelligent matching of GPUs and NICs, automatic recognition of physical topologies inside and outside nodes, and topology-aware congestion-free communication algorithms through the ACCL communication library, completely eliminating network congestion, improving network communication efficiency, and enhancing the scalability of distributed training systems. It can achieve more than 80% linear cluster performance at the scale of 10,000 cards, and more than 95% effective (computing) performance at the scale of 100 cards, which can meet the needs of more than 80% of business scenarios.
3. High-performance Active Data Loading Acceleration Software KSpeed. PAI Lingjun developed the high-performance active data loading acceleration software KSpeed based on high-performance network RDMA and high-performance communication ACCL to conduct intelligent data IO optimization. The compute-storage separation architecture is widely used in AI, HPC and big data business scenarios, but the loading of a large number of training data easily forms an efficiency bottleneck. Alibaba Cloud uses the high-performance active data loading acceleration software KSpeed to achieve an order-of-magnitude improvement in data IO performance.
4. GPU Container Virtualization Solution eGPU. Aiming at the practical problems encountered in business scenarios such as large-scale AI jobs, expensive GPU hardware resources, and low cluster GPU utilization rate, PAI Lingjun supports the GPU virtualization technology eGPU, which can effectively improve the GPU utilization rate of AI clusters. The specific advantages are as follows: Supports free splitting in two dimensions of video memory and computing power. Supports multiple specifications. Supports dynamic creation and destruction. Supports hot upgrade. Supports user-mode technology to ensure higher reliability.

搜集汇总

背景与挑战

背景概述

PAI灵骏智算服务提供高性能异构算力与智能调度能力，支持万级GPU弹性扩展和微秒级低时延，显著提升资源利用率和训练效率。该服务适用于AI研发、自动驾驶及科学计算等场景，具备RDMA网络架构、虚拟化技术和智能数据加速等核心功能。