Performance profile configuration

Caution

Performance profile installation and accelerator configuration is supported exclusively for Intel Gaudi accelerators present in Ubuntu clusters.

Performance profiles enable you to optimize system performance for specific workloads. Omnia supports the configuration of performance profiles for Ubuntu clusters that have Intel Gaudi accelerators. These profiles come with predefined settings tailored to different use cases. For more information, click here.

Prerequisites

• Create an Inventory file

To configure performance profiles, list all the nodes for which you want to apply the profiles in an inventory file. A sample inventory looks like:

node3
node1

• Configure Performance profiles

In the utils/performance_profile/performance_profile_config.yml file, under intel_gpu, add or alter the values based on the following list of parameters:

Parameters for performance profile configuration

Parameters	Details
performance_profile String Required	The field captures the performance profile which will be configured on the nodes. To see and get a brief summary of all of the available profiles, use the tuned-adm list command. Example: accelerator-performance
performance_profile_plugin String Optional	If you want to customise the default performance profiles, you can do so by adding multiple plugins/parameters or by altering the default values present here. If there is no need to modify the profile, the performance_profile_plugin section can be left as it is. By default, it takes the pre-defined plugin values placed in the profile. Example: If you have selected accelerator-performance as your performance profile, then you can alter the values of parameters such as force_latency and vm.swappiness. If required, you can also add supported other supported parameters such as vm.nr_hugepages with their desired values.
reboot_required String Required	The value provided for this field denotes if the device should be rebooted after peformance profile configuration. Choices: yes no -> default

Here’s a sample of the default performance_profile_config.yml file:

intel_gpu:
  performance_profile: "accelerator-performance"
  performance_profile_plugin:
    sysctl:
      - vm.nr_hugepages: 156300
  reboot_required: "no"

Here’s an example for adding/modifying multiple plugins in the performance_profile_config.yml file:

intel_gpu:
performance_profile: "accelerator-performance"
performance_profile_plugin:
  sysctl:
    - vm.nr_hugepages: 156300
  cpu:
    - force_latency: 99
  disk:
    - read_ahead_kb: 4096
  reboot_required: "no"

Note

For Intel Gaudi accelerators, Omnia recommends to add the vm.nr_hugepages as a profile parameter under sysctl plugin and set its value to 156300.

• Execute the the playbook

Run the playbook using the following commands:

cd utils/performance_profile
ansible-playbook performance_profile.yml -i inventory

If you have any feedback about Omnia documentation, please reach out at omnia.readme@dell.com.