Loading DPL Applications

NVIDIA DPL programs are deployed to the NVIDIA® BlueField® networking platform (DPU or SuperNIC) using the P4Runtime API. This allows platform-independent, standards-based integration with SDN controllers.

The DPL compiler generates a pipeline binary optimized for execution on BlueField. Pipeline loading and control—such as installing the program and populating P4 tables—are handled via the P4Runtime API, an open and well-defined interface.

The P4Runtime server, running on the BlueField device, enables a P4 Controller to:

• Connect over gRPC

• Set the ForwardingPipelineConfig (i.e., install the compiled DPL binary and p4info)

• Query the device for its current pipeline config and table state

• Maintain runtime P4 tables as defined in the DPL source

This model enables integration with open-source, proprietary, or custom-built controllers in a standardized way.

Before loading a DPL application, ensure the following services and components are properly set up:

• DPL Runtime Service is running and configured on the BlueField (Arm side). See the DPL Container Deployment page for setup instructions.

• DPL Development Container and the p4runtime_sh.sh launch script are installed on the host. See the DPL Installation Guide for more details.

A DPL application can be loaded using:

• A custom P4Runtime controller

• The NVIDIA-supplied Python controller

• An open source controller

In the following example, we'll be using the P4Runtime controller bundled within the DPL Development Container.

Using p4runtime_sh.sh Script

            

            
usage: ./p4runtime_sh.sh [help] [-i <docker image>] [-a <dpl_rtd_ip:port>] -p <program_folder> -d <device ID> [OPTIONS]
For example:
./p4runtime_sh.sh -p /root/hello_packet/_out -a 192.168.1.100:9559 -d 1000
        

Arguments:

• -i - The pulled DPL Development Container image.

• -p - Directory that holds the DPL program compilation outputs (Compiling DPL Applications).

• -a - Address of the DPL Runtime daemon and the P4Runtime port as specified in the DPL Runtime DPL Service Configuration.

• -d – ID of the target device

• --mount - (Optional) Additional volume to mount to the container. Can be used multiple times.

Example Command

            

            
./p4runtime_sh.sh -d 1000 -p /root/hello_packet/_out
        

p4runtime_sh python application

After successfully loading a program, the script launches an interactive Python shell (p4runtime_sh python application) connected to the P4Runtime server on the dpl-rt-service running on the DPU.

From here, you can inspect and manage tables entries using this tool. For more details about using the p4runtime_sh application see:

• Our P4 Runtime Controller documentation.

• Official p4runtime-shell project documentation.

DPL table entries are added via a P4Runtime controller, which may run remotely or locally on BlueField.

• The standard gRPC-based controller model supports ~50K rule insertions per second

• For use cases requiring high-speed (1M+) rule insertions, NVIDIA is introducing a bulk insertion API extension to the P4Runtime protobuf specification

  Note

  This feature is planned for future releases.

Loading a program might fail when the dpl-rt-service processes the compiled program due to various reasons (mostly configuration issues), in such cases, you might see error message like:

            

            
# ./scripts/p4runtime_sh.sh --device-id 1000 -p tables/_out/
CRITICAL:root:Error when setting config
CRITICAL:root:P4Runtime RPC error (INVALID_ARGUMENT): Error setting program: Failed creating P4 tables
        

Refer to the following for troubleshooting common issue:

            

            
# ./scripts/dpl_admin.sh config
        

            

            
# ./scripts/dpl_admin.sh nv-config