Skip to content

[FEA] Add a DAQIRI raw Ethernet to UCX GPU egress example #288

Description

@dleshchev

Is your feature request related to a problem? Please describe.

DAQIRI users need a practical example of receiving raw Ethernet data, processing it with CUDA, and forwarding the result to another GPU-capable system using UCX/UCP over RoCEv2.

Describe the solution you'd like

Add an opt-in two-host example implementing:

Raw data source → DAQIRI raw RX → batched CUDA processing
                → UCX/RoCEv2 → GPU-accessible receiver

The example should demonstrate:

  • Receiving and assembling packetized data with DAQIRI.
  • Applying representative batched CUDA processing.
  • Sending processed payloads through UCX/UCP.
  • Receiving into GPU-accessible memory.
  • Bounded receiver capacity and overload accounting.
  • UCX-only, staged, and complete-pipeline benchmarks.
  • A UCX-enabled container, example configurations, and a tutorial modeled after the ResNet50 example.
  • Mapped pinned memory support for DGX Spark, while retaining CUDA-device support for other systems.

Describe alternatives you've considered

Using DAQIRI multi-engine support, a DAQIRI UCX engine, custom verbs egress, or host-staged sockets would add unnecessary complexity for an example application.

Additional context

A prototype is available on feat/ucx-gpu-egress-example.

On the two-Spark setup, it delivered and GPU-validated 1.6 million processed payloads at approximately 94.94 Gbit/s, with zero drops, sequence gaps, or validation failures.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions