AI 模型和平台
Oracle Acceleron 新增前端 RDMA和多Planar Felover 用于AI
在 Google 上的首选来源中添加 Unite.AI
Oracle 宣布新建的Oracle 2026年10月27日在拉斯维加斯的Oracle AI World提供AI工作量的Aceleron云联网能力，引入了前端RDMA连接，扩展了NVMe块存储选项，Oracle 称改进了高达3x的性能，为其多行星网络建立了一个集体通信库。
兹Oracle Acceleron是Oracle Cloud基础设施的一套网络和基础设施技术，它结合了专用的网络织物，更直接的数据路径，以及一个趋同的NIC，与主机级的零信任路由和多行星设计，以保障安全和恢复能力。 新能力跨越三个领域：AI数据移动，分布AI通信，以及安全的多租户AI基础设施。
Oracle的公告中，称 Pradep Vincent是计算和AI基础设施的执行副总裁。 Vincent 称公司正在建造Acceleron以响应，方法是改变交通路线，从故障中恢复过来，以及帮助客户的尺度安全。
前端 RDMA 和 NVMe 存储器
第一个新能力是Oracle Acceleron高性能Fabric，带来 RDMA连接到OCI前端计算，Oracle 称它帮助客户部署推论，存储，以及HPC工作量，性能更高。 第二，扩展的NVMe块存储附件选项横跨Oracle Acceleron裸金属和虚拟机形状，根据公告，提升了最多3x的AI，数据库等数据密集型工作量的性能。
在Oracle的OCI博客上的技术文章中，作者描述RDMA是在CPU参与最小的不同机器上的内存之间直接移动数据，绕过常规传输中的大部分操作系统和软件处理以减少延迟，维持高吞吐量，并留下更多的CPU能力用于应用程序。 作者写道，这些好处超越了GPU到GPU的通信到存储访问和跨集群传输，但RDMA对一个共享的混合工作负荷网络的拥堵，耐久和包丢失非常敏感。
一个公平问题是，RDMA通常使用基于速率的拥堵控制，而TCP则使用基于窗口的控制，根据流量模式，一个循环可以压倒另一个循环。 为了维护公平，该帖子称Oracle Acceleron SmartNICs使用等级标志桶来跟踪RDMA和非RDMA流量的带宽消耗：一个母桶代表每个流量等级与儿童桶共享带宽，并作为标志耗尽SmartNIC概率标记ECN能力包或无法响应的投放包，产生逐渐拥堵信号。 这个机制是工作节约的，因此任一交通类都可以借用未使用的带宽，它也处理一个顶端开关连接到SmartNIC的插播。
对于织物内部的拥堵，OCI在Acceleron SmartNICs中使用单向延迟拥堵控制。 测量前进路径的延迟可以避免虚假的拥堵信号，当只有逆向路径被拥堵时，作者们解释道。 当一流的显示提升延迟时，SmartNIC探测到三条交替路径，如果它发现是一个更好的路径，则移动流量；否则它通过ECN标记或包滴来信号拥堵。 OCI的前端RDMA支持所有队列对称类型和动词，允许客户运行现有的应用程序而无需修改。
集体通信和多计划故障
该公告的分布式通信能力是Oracle集体通信图书馆，即OCCL，以及透明的前端多行星网络。 Oracle 称 OCCL 帮助在网络路径失败时维持分布式AI工作量运行，方法是将集体流量分布到OCI的后端多Planar Network上，引导流量向更健康的路径发展，并在链接，端口或NIC失败时恢复。 透明前端多Planar Networking 带来多Planar 网络到前端网络，内置透明路径选择和故障，外加透明前端RDMA流量拥堵控制。
博客在训练术语中将问题设定为：同步的工作移动速度只和链条中最慢的环节一样快，因此单一的拥堵或失败路径可以产生全工作量的减速，标准等价的多路径转发针头流到路径时不知道其大小，产生不均匀的链接利用率。 多行星织物也要求具有专门功能的多港NIC，作者写道这一点并不总是可行的；OCI建造OCCL是为了消除对一般AI和机器学习工作量的这种限制。
OCCL是常用的集体通信库的插件，AI框架用来协调GPU之间的数据交换。 根据该帖子，它通过多个队列对流分配流量，在队列对流经历较高的吞吐量时发送更多的流量，并积极将流量移动到不太拥挤的路径，通过标准的CCL网络-插座接口执行拥堵-意识到消息级别和块级负载平衡。 它根据观测到的拥堵和健康状况，动态地分配跨标准可靠连接队列的集体通信信息，并支持端口级和NIC级故障，在链接或NIC故障后引导集体通信，因此分布式培训没有重新开始。
作者们写道，将故障转换、恢复和拥塞感应负载平衡到CCL层中，将这些功能从网络硬件和固件中脱钩，并在单板和多板网络上工作，他们补充说，客户不需要复杂的路由控制器来选择路径或检测路径的变化。 该帖子日期为2026年10月27日，由OCI Networking Ops的副总裁兼杰出工程师贾格·布拉尔(Jag Brar)，文森特(Vincent)和OCI建筑师乔伊·甘古利(Joy Ganguly)撰写。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
AI Models & Platforms
Oracle Acceleron Adds Front-End RDMA and Multiplanar Failover for AI
Add Unite.AI to your preferred sources on Google
Oracle announced new Oracle Acceleron cloud networking capabilities for AI workloads at Oracle AI World in Las Vegas on October 27, 2026, introducing front-end RDMA connectivity, expanded NVMe block-storage options that Oracle said improve performance up to 3x, and a collective communications library for its multi-planar networks.
Oracle Acceleron is Oracle Cloud Infrastructure’s suite of networking and infrastructure technologies for running demanding AI workloads reliably, securely, and at scale. It combines dedicated network fabrics, more direct data paths, and a converged NIC, with host-level zero-trust routing and multi-planar designs for security and resilience. The new capabilities span three areas: AI data movement, distributed AI communication, and secure multi-tenant AI infrastructure.
“When thousands of accelerators are working together, one congested or failed path can have consequences far beyond a single connection,” said Pradeep Vincent, executive vice president of Compute and AI Infrastructure at OCI, in Oracle’s announcement. Vincent said the company is building Acceleron to respond by rerouting traffic, recovering from failures, and helping customers scale securely.
Front-End RDMA and NVMe Storage
The first new capability, Oracle Acceleron High-Performance Fabric, brings RDMA connectivity to OCI front-end compute, and Oracle said it helps customers deploy inference, storage, and HPC workloads with greater performance. The second, expanded NVMe block-storage attachment options across Oracle Acceleron bare metal and virtual machine shapes, improves performance up to 3x for AI, database, and other data-intensive workloads, according to the announcement.
In a technical post on Oracle’s OCI blog, the authors described RDMA as moving data directly between memory on different machines with minimal CPU involvement, bypassing much of the operating-system and software processing in a conventional transfer to reduce latency, sustain high throughput, and leave more CPU capacity for applications. Those benefits extend beyond GPU-to-GPU communication to storage access and cross-cluster transfers, the authors wrote, but RDMA is highly sensitive to congestion, latency, and packet loss across a shared, mixed-workload network.
One fairness problem is that RDMA typically uses rate-based congestion control while TCP uses window-based control, and depending on traffic patterns one loop can overwhelm the other. To preserve fairness, the post states that Oracle Acceleron SmartNICs use hierarchical token buckets to track bandwidth consumption for RDMA and non-RDMA traffic: a parent bucket represents shared bandwidth with child buckets for each traffic class, and as tokens deplete the SmartNIC probabilistically marks ECN-capable packets or drops packets that cannot respond, producing a gradual congestion signal. The mechanism is work-conserving, so either traffic class can borrow unused bandwidth, and it also addresses incast where a top-of-rack switch connects to a SmartNIC.
For congestion inside the fabric, OCI uses one-way-delay congestion control in the Acceleron SmartNICs. Measuring forward-path delay avoids the false congestion signal a round-trip measurement can produce when only the reverse path is congested, the authors explained. When a flow shows elevated delay, the SmartNIC probes three alternate paths and moves the flow if it finds a better one; otherwise it signals congestion through ECN marking or packet drops. OCI’s front-end RDMA supports all queue pair types and verbs, allowing customers to run existing applications without changes.
Collective Communications and Multi-Planar Failover
The announcement’s distributed-communication capabilities are the Oracle Collective Communications Library, or OCCL, and Transparent Front-End Multiplanar Networking. Oracle said OCCL helps keep distributed AI workloads running when network paths fail by spreading collective traffic across OCI’s backend Multiplanar Network, steering traffic toward healthier paths and recovering when links, ports, or NICs fail. Transparent Front-End Multiplanar Networking brings multiplanar networking to front-end networks with built-in transparent path selection and failover, plus transparent congestion control for front-end RDMA traffic.
The blog frames the problem in training terms: synchronized jobs move only as fast as the slowest link in the chain, so a single congested or failed path can produce a workload-wide slowdown, and standard equal-cost multi-path forwarding pins flows to paths without knowing their sizes, producing uneven link utilization. Multi-planar fabrics have also required multi-port NICs with specialized functionality, which the authors wrote is not always feasible; OCI built OCCL to remove that constraint for general AI and machine learning workloads.
OCCL is a plugin for commonly adopted collective-communication libraries, which AI frameworks use to coordinate data exchange among GPUs. According to the post, it distributes traffic across multiple queue pairs, sends more traffic over queue pairs experiencing higher throughput, and actively moves traffic to less congested paths, implementing congestion-aware message-level and chunk-level load balancing through the standard CCL network-plugin interface. It dynamically assigns collective-communication messages across standard Reliable Connection queue pairs based on observed congestion and health, and it supports port-level and NIC-level failover, steering collective communication after a link or NIC failure so distributed training continues without a job restart.
Moving failover, recovery, and congestion-aware load balancing into the CCL layer decouples those functions from network hardware and firmware and works on both single-planar and multi-planar networks, the authors wrote, adding that customers do not need complex routing controllers to select paths or detect path changes. The post, dated October 27, 2026, was written by Jag Brar, vice president and distinguished engineer for OCI Networking Ops; Vincent; and Joy Ganguly, an OCI architect.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.