This article was accepted into the corpus but its outbound wikilinks were never NER-processed — typical at the deepest BFS hop or when the run's entity cap was reached. No expansion funnel to show.
| RDMA over Converged Ethernet | |
|---|---|
| Name | RDMA over Converged Ethernet |
| Abbreviation | RoCE |
| Introduced | 2010s |
| Type | Networking protocol |
| Developer | InfiniBand Trade Association; Mellanox Technologies; Broadcom |
RDMA over Converged Ethernet RDMA over Converged Ethernet enables low-latency, high-throughput remote direct memory access across Ethernet networks, combining elements from InfiniBand, Ethernet, Internet Engineering Task Force, Mellanox Technologies, and Broadcom. The technology arose from collaborations among the InfiniBand Trade Association, Microsoft, Facebook, Amazon (company), Google, and Intel Corporation to extend memory semantics to modern datacenter fabrics and to support workloads from Hadoop to VMware ESXi and Kubernetes.
RoCE unites design ideas from InfiniBand with the ubiquity of Ethernet and the governance of the Institute of Electrical and Electronics Engineers to provide kernel-bypass, zero-copy transfers for NVIDIA GPU-accelerated workloads, Supermicro server deployments, and hyperscale operators like Microsoft Azure and Google Cloud Platform. Vendors such as Mellanox Technologies, Broadcom, Intel Corporation, Cisco Systems, and Arista Networks implement RoCE in converged top-of-rack and leaf-spine architectures used by organizations including Facebook, Amazon Web Services, and Oracle Corporation.
RoCE exists in versions rooted in protocols standardized by groups like the IETF and design contributions from the InfiniBand Trade Association; these include variants aligned to UDP and Ethernet II framing and that map InfiniBand verbs over Layer 2 or Layer 3 fabrics. Key protocol components reference the RDMA verb semantics pioneered in InfiniBand, link management influenced by IEEE 802.1Q, and congestion control techniques derived from research by Microsoft Research, Stanford University, University of California, Berkeley, and Rice University. Implementations often incorporate congestion control algorithms such as DCQCN and TIMELY researched at Facebook and Google, and leverage offloads akin to those in TCP Offload Engine products from Solarflare and Chelsio Communications.
Hardware ecosystems for RoCE include NICs from Mellanox Technologies (now part of NVIDIA), Intel Corporation Ethernet controllers, and ASICs from Broadcom and Marvell Technology Group, deployed in switches by Cisco Systems, Arista Networks, and Juniper Networks. Server vendors like Dell Technologies, Hewlett Packard Enterprise, and Lenovo offer RoCE-capable systems integrated with accelerators from NVIDIA and AMD. Firmware, drivers, and firmware management are provided by projects and vendors including OpenFabrics Alliance, Linux kernel, Microsoft Windows Server, and VMware ESXi. Storage arrays from NetApp and Pure Storage and software-defined storage platforms like Ceph and GlusterFS also integrate RoCE-enabled paths.
When properly configured, RoCE provides microsecond-scale latency competitive with InfiniBand and throughput approaching line-rate on 25/40/100/200/400 Gigabit Ethernet fabrics used by HPC centers like Oak Ridge National Laboratory and cloud providers such as Amazon Web Services. Scalability depends on fabric topologies used by Facebook and Google, congestion control from Microsoft Research and Stanford University, and features like Priority Flow Control from IEEE 802.1 committees. Benchmarking often references tools and suites from SPEC, Intel MPI, OpenMPI, and vendor benchmarks by NVIDIA Mellanox and Cisco Systems.
RoCE supports high-performance computing deployments at facilities like Argonne National Laboratory, low-latency financial trading platforms in infrastructures used by CME Group traders, distributed machine learning frameworks such as TensorFlow, PyTorch, and Horovod, and storage protocols including NVMe over Fabrics and block services used by VMware vSphere. Cloud-scale services at Microsoft Azure, Google Cloud Platform, and Amazon Web Services use RoCE for virtualization, container orchestration with Kubernetes, and persistent storage for systems developed by OpenStack and the Linux Foundation.
Standards and interoperability efforts involve the InfiniBand Trade Association, IEEE, IETF, and the OpenFabrics Alliance to harmonize RoCE with Ethernet features like IEEE 802.1Q tagging, Data Center Bridging extensions, and Priority Flow Control. Vendor interoperability testing occurs at industry events hosted by Interop, Open Compute Project, and consortia including SNIA and the Storage Networking Industry Association. Compatibility matrices encompass firmware and driver stacks from NVIDIA, Intel, Broadcom, Mellanox Technologies, and operating systems such as Linux kernel and Microsoft Windows Server.
Security and management practices for RoCE align with orchestration and monitoring platforms like Prometheus, Grafana, and Ansible and with access controls used in environments managed by Red Hat and Canonical. Threat models reference guidance from NIST and integrate with enterprise networking controls from Cisco Systems and Juniper Networks; management workflows use telemetry standards promoted by OpenTelemetry and configuration models from NETCONF and RESTCONF. Network administrators deploy congestion and flow control policies derived from IEEE 802.1 specifications and coordinate with cloud operators at Microsoft and Google to maintain service-level objectives.
Category:Computer networking