Is CuBB E2E test in one GH200 machine possible?

Accelerated Computing Aerial Forum
1

Hi, I have a question regarding the setup described in the documentation.

The guide shows that the gNB should be run on a GH machine, and the RU emulator should be run on a separate R750 machine, then both are connected to perform cuBB testing.

However, I would like to know if it’s possible to run both the gNB and the RU emulator on a single GH machine for testing purposes.

Is this setup supported, and are there any known limitations or configuration changes required to do so?

Thanks!

2

It is doable but hasn’t been tested yet. will get you back for the setup recommendations.

3

Thank you for your response!

I will try it on my side as well while waiting for your setup recommendations.

4

@choss2

You can test cuBB E2E on a single MGX GH with assumption that you have two BF3 card installed.

  • setup: one way, you can connect one port of each BF3 (e.g. BF3_1: P0/aerial00, and BF3_2: P0/aerial02)) to switch. Another way, you could do direct connection from one port of the 1st BF3 to one port of the 2nd BF3.

  • PTP configuration: need one ptp.conf, one ptp4l.service and two phc2sys.services.
    For the ptp.conf and ptp4l.service configuration, please refer to the guidance Installing Tools on Grace Hopper MGX System — Aerial CUDA-Accelerated RAN .
    For the phc2sys, services, the phc2sys.service, use
    ExecStart=/bin/sh -c “/usr/sbin/phc2sys -s /dev/ptp$(ethtool -T aerial00 | grep PTP | awk ‘{print $4}’) -c CLOCK_REALTIME -n 24 -O 0 -R 256 -u 256”`

    2nd phc2sys2.service, use
    ExecStart=/bin/sh -c “/usr/sbin/phc2sys -c /dev/ptp$(ethtool -T aerial02 | grep PTP | awk ‘{print $4}’) -s CLOCK_REALTIME -n 24 -O 0 -R 256 -u 256”`

  • At last, please make sure the cpu affinities in cuphycontroller_F08_CG1.yaml, l2_adapter_F08_CG1.yaml, test_mac_config and ru_emulator/config.yaml, are not overlapped.

Please try it out and let us know if your have any problems.

Thanks!

5

Thank you for the detailed instructions!

Our lab is planning to purchase the necessary hardware to build a full Aerial platform later this year.
However, before the equipment arrives, I am currently testing Aerial using a Grace Hopper-based GH instance provided by Lambda’s cloud platform.

Since I don’t have control over the hardware in this cloud environment, I’m unable to install BF3 cards directly.
Given this limitation, would it be difficult or impossible to perform an end-to-end test that connects the RU emulator and gNB on such a GH cloud instance?
Or, alternatively, what level of testing is feasible on a GH-based Lambda cloud instance?

I’d appreciate your guidance on what’s realistically achievable in this setup.

Thanks again!

6

Hi @choss2

Thank you very much for sharing your plan about building a full Aerial platform.
For the current accessible HW which is GH without FH NIC, it is not possible to run cuBB e23 test, but you can run cuPHY Module tests. Please refer to the guidance here
Running Aerial cuPHY — Aerial CUDA-Accelerated RAN

7

Hi, I understand that on the Lambda GH cloud environment, only the Aerial cuPHY module tests are currently feasible due to the lack of a BF NIC.

I’m wondering if it’s possible to integrate this cuPHY module test setup with an external OAI-based gNB/UE simulator to enable user application(e.g., cloud gaming, web loading) performance testing.

Would such integration be feasible in the current Aerial architecture?

If such integration is not supported,
does NVIDIA provide any subscription-based access to an Aerial-enabled cloud instance (e.g., with GH200 and BF3 NIC) that allows for full-stack end-to-end testing, including RU↔DU↔CU↔UE application flow?

Thanks again for your guidance!