Networks & InfrastructureOracleOcclAI TrainingRdma

Oracle's OCCL plugin lifts AI training bandwidth by up to 14%

Oracle's OCCL network plugin, preinstalled on OCI HPC images, lifted NCCL bus bandwidth up to 10% for AllReduce and over 14% for AllToAll in internal GB300 benchmarks.

3 min read

Oracle bolsters cloud plumbing for distributed AI trainingNetworks & Infrastructure
Oracle bolsters cloud plumbing for distributed AI training12019 / pixabay

Why it matters

  • Oracle's OCCL plugin raised AllToAll bandwidth by over 14% and AllReduce by 2-10% in internal benchmarks on Nvidia GB300 systems.
  • Benchmarks used clusters of 16 to 128 nodes with four GPUs per node.
  • OCCL supports RoCEv2 and InfiniBand and integrates with Nvidia NCCL and AMD RCCL via standard network-plugin interfaces.
  • The plugin is preinstalled in supported Oracle HPC images on OCI, requiring no separate control plane or agents.
  • OpenAI's MRC protocol, backed by AMD, Broadcom, Intel, Microsoft and Nvidia, runs on all OpenAI supercomputers including the Oracle-built Zettascale10.

The story

Oracle has recorded a 14% average bandwidth gain for AllToAll operations with its new network plugin for distributed AI training, according to internal benchmarks run on Nvidia GB300 systems. The tool, called the Oracle Collective Communications Library (OCCL), is now preinstalled in supported Oracle HPC images within Oracle Cloud Infrastructure, with no separate control plane, agents, or application-side services required.

For operators and cloud teams running large GPU clusters, the significance is straightforward. Distributed training jobs fail or slow down when network paths congest, or when links and network interface cards drop. OCCL addresses both failure modes in software, rather than through added hardware.

How does OCCL fit into existing stacks?

OCCL integrates with Nvidia's NCCL and AMD's RCCL collective-communication libraries through their standard network-plugin interfaces. AI frameworks continue to use NCCL or RCCL unchanged. The plugin intercepts inter-node data transfers through the standard interface and manages the network paths those transfers use.

The software supports RoCEv2 and InfiniBand transports. It monitors queue-pair transfers and dynamically distributes traffic across network paths showing less congestion. When a supported link or NIC fails, OCCL automatically reroutes traffic around the failure and restores connectivity once the hardware recovers. It handles single-planar and multi-planar RDMA networks in a vendor-agnostic fashion, which Oracle says preserves flexibility in the underlying network hardware.

What do the benchmarks show?

Oracle's internal tests used Nvidia GB300 systems with clusters of 16 to 128 nodes, each node carrying four GPUs. For AllReduce, the collective operation that aggregates and synchronizes data across every worker GPU, OCCL increased average NCCL bus bandwidth by 2% to 10%. For AllToAll operations, where every worker node or GPU exchanges data pieces with every other node in the cluster, the gain came in at just over 14%.

Those are modest but meaningful numbers for training economics. A 10% bandwidth improvement on an AllReduce-heavy job translates directly into shorter synchronization stalls, and less idle GPU time, across a cluster that may run for weeks.

Why multi-planar networks matter here

Lilian Quan, senior principal solutions architect for Oracle, explained the design rationale: "OCI GPU clusters use RDMA networks that can be deployed with a multi-planar architecture, in which network capacity is provided through multiple independent planes."

"This design creates path diversity and helps isolate failures," Quan said. "To gain the full benefit, collective traffic must use capacity across the available planes and respond to congestion or failures without interrupting the training job."

The plugin places congestion control and link/NIC failure recovery intelligence in the collective-communication transport layer. That placement, rather than any single feature, is Oracle's core argument.

Where does this sit in the broader standards picture?

OCCL arrives amid wider industry effort to harden networks for AI workloads. OpenAI, working with AMD, Broadcom, Intel, Microsoft, and Nvidia, developed the Ethernet-based Multipath Reliable Connection (MRC) transport protocol. MRC is currently deployed across all of OpenAI's largest supercomputers, including the Oracle-built Zettascale10, as reported in the latest SDxCentral Magazine.

The contrast is worth noting for practitioners. MRC requires protocol-level change in the Ethernet transport. OCCL works within the existing NCCL and RCCL plugin interfaces on OCI today. Oracle has positioned the plugin as a software-defined route to resilience and traffic distribution, keeping hardware choices open.

The benchmarks cited are Oracle's own internal figures, not independent third-party results, so buyers should treat the 2% to 14% range as vendor-reported until verified on their own workloads. For OCI HPC customers, the plugin needs no installation effort, and its failure-rerouting behavior applies immediately to supported configurations.

Also reported

Share this article:

« PreviousNext »

More from Daniel Okafor

Daniel Okafor

Show full bio

Senior reporter covering regulation, policy and their impact on Carrier business, 5G and 6G, fiber and broadband at Telecom Gazette.

536 articles