gh200-Ubuntu-22.04-autoinstall

Automated installation of GH200 system using Ubuntu 22.04 using USB drive

A lot of this came from the Official Nvidia Ubuntu 22.04 Grace Installation Guide. If you have problems or want to install manually, please see that guide for more details. There is also a Grace Performance Tuning Guide that may be helpful.

This repo was created based on testing of the GH200 system (specifically the Supermicro ARS-111GL-NHR). The systems I'm testing on also have a single Bluiefield-3 installed (although 2 BF-3 have also been tested), but this should not matter for the installation.

The contained scripts will perform the following actions:

During Installation:

• Create a new user (as defined by user)
• Make linux-nvidia-64k-hwe-22.04 the default kernel
• Install the Nvidia MLNX-OFED repo
• Install the Nvidia CUDA SBSA repo
• Update/Upgrade all packages
• Install the First-boot service that will run on the first boot

(NOTE: Nvidia requires MLNX-OFED to be installed BEFORE the GPU drivers for NCCL to work correctly)

During First Boot of system

• Install mlnx-fw-updater mlnx-ofed-all
• Install cuda-drivers and nvidia-kernel-open drivers
• Install cuda-toolkit-12-4 nvidia-container-toolkit
• Update and enable the nvidia-persistenced service due to bug
• Disable irqbalance service (recommended by performance tuning guide)
• Disable NUMA balancing (recommended by performance tuning guide)
• Automatically load the nvidia-peermem kernel module (required for NCCL and GPUDirect)
• Disable the first-boot service
• Reboot the system

Creating Installable USB drive via Rufus

There are multiple ways to create an installable Ubuntu 22.04 USB drive. I used Rufus on a Windows-11 system.

1. Download Rufus Portable and the latest Ubuntu-22.04.4 ISO.
2. Select USB drive.
3. Install as ISO (enuring you can modify the files on the USB afterwards)

Customize the user-data file

Before copying over the files, you'll need to customize the cidata/user-data file with your installation details

Updating with CIDATA for auto-install

After creating a bootable Ubuntu installation drive, copy the files from cidata to the

1. Create directory cidata in the root of the Ubuntu USB drive

2. Copy All files over to the cidata directory on the Ubuntu USB drive

   • user-data : Ubuntu Autoinstall file
   • meta-data : Ubuntu Meta-data files
   • first-boot.service : One-shot service to launch the first-boot.sh script
   • first-boot.sh : Script that will run on first boot

3. Update the boot/grub/grub.cfg file and add the following menuentry to the list:

menuentry "Install GH200 System (Requires Internet)" {
        set gfxpayload=keep
        linux   /casper/vmlinuz quiet autoinstall 'ds=nocloud-net;s=file:///cdrom/cidata/'
        initrd  /casper/initrd
}

Install the system via USB drive

NOTE: There is a bug with the Supermicro ARS-111GL-NHR that causes the system to hang when rebooting with a USB drive connected. Unknown if this affects all GH200 systems. After installation, it's likely the screen is stuck. Simply remove the USB drive (or power cycle from BMC) and the system should boot right into the OS.

Validating everything is working

Run the following commands to validate all drivers are working:

1. Confirm the nvidia drivers have loaded correctly:

   
   nvidia-smi

+-----------------------------------------------------------------------------------------+ | NVIDIA-SMI 555.42.06 Driver Version: 555.42.06 CUDA Version: 12.5 | |-----------------------------------------+------------------------+----------------------+ | GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | | | | MIG M. | |=========================================+========================+======================| | 0 NVIDIA GH200 480GB On | 00000009:01:00.0 Off | 0 | | N/A 33C P0 76W / 900W | 1MiB / 97871MiB | 0% Default | | | | Disabled | +-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+ | Processes: | | GPU GI CI PID Type Process name GPU Memory | | ID ID Usage | |=========================================================================================| | No running processes found | +-----------------------------------------------------------------------------------------+

2. Ensure the nvidia-peermem kernel module has loaded correctly

lsmod | grep nvidia_peermem

3.