doca-urom-svc

What it does

Operates the DOCA UROM Service container on BlueField Arm to execute remote memory operations (puts, gets, atomics, collectives) enqueued by a paired host using the doca-urom library. It pulls an NGC image, selects the UCX component, sizes queues, configures Comch pairing, and aligns host and service versions. It enforces security boundaries by relying on Comch pairing and RDMA permissions, and instructs to pair only intended hosts, expose least-privilege memory regions, and verify both views before start. It triggers for slow UCX collectives, unexpected NOT_PERMITTED, or missing completions. It does not apply to host application code, MPI/UCX integration design, or DOCA install.

How it works

The skill describes operating steps for deploying and configuring the service container: pull the image from NVIDIA NGC, choose UCX components and collectives to expose, determine enqueue queue depths, set up Comch pairing with the host library, and ensure host/service version compatibility. It emphasizes pre-start verification of Comch and RDMA permissions, and ongoing monitoring of service logs and RDMA counters to confirm correct operation. It defines error taxonomy around NOT_PERMITTED and completions, and guides debugging by separating service queue issues from substrate transport problems, then evaluating offload performance against host baseline as part of safety policy.

When to use it

Load this skill for hands-on DOCA UROM Service deployment work on BlueField where DOCA is already installed. Use when deciding if DOCA UROM Service is appropriate, deploying the service container, choosing configuration axes, pairing host and service, verifying the paired contract, and validating logs and RDMA counters before scaling workloads. It is not intended for general DOCA orientation, DOCA installation, host-side library API questions, or non-UROM HPC topics.

What it can touch

The skill references interacting with the DOCA UROM service container, the host doca-urom library, the doca-rdma transport substrate, and the published configuration surface (flags/env like SERVICE_ARGS, UROM_PLUGIN_PATH, and mounting plugins/). It also mentions observing the service logs and RDMA counters.

Caveats

License: Apache-2.0. It requires BlueField-Arm-only DOCA service container pulled from NVIDIA NGC and started under the BlueField OS container runtime; host-side install is irrelevant. There is no service-side host-endpoint authorization list; access is governed by Comch pairing and underlying RDMA permissions. The skill explicitly warns not to load for host-side library APIs, or pre-baked production bundles, and cautions about version pairing between host library and service.