Support GPU types

[a-thaler]

Description

As Sebastian and Matthew, we would like to be able to run GPU-intensive workloads on Kyma Runtime so that we can build and deploy machine learning models or data analytics tasks efficiently. That story is part of https://github.tools.sap/kyma/backlog/issues/4633

Allowing GPU workloads would make Kyma Runtime a more versatile and competitive product, attracting users who require heavy computational tasks, including but not limited to data scientists, ML engineers, and computational physicists.

Acceptance Criteria

1. Users should be able to specify GPU resources in their Kyma Runtime configurations.
2. The system should schedule the workload on nodes with available GPU resources.
3. Performance metrics specific to GPU should be available for monitoring.
4. Must comply with security and resource isolation standards.
5. Additional infrastructure setup like drivers provisioned via a DaemonSet should be provided in a convenient way

Non-functional Requirements

1. Scalability: Should support clusters with multiple GPUs.
2. Security: Isolation of GPU resources to prevent unauthorized access.
3. Performance: Minimal latency when scheduling and running GPU workloads.

Useful Links

1. Kubernetes GPU support documentation

Tasks

Preparation

• Agree on the concrete types and CU calculation, check regional availability
• Come up with a concrete concept by investigating what else is needed besides the actual machine type Implementation: [PoC] Run GPU workload in Gardener cluster and provide concept how to enable GPU in Kyma Runtime #18771
• Do extensive tests for all modules having daemonsets
• Enable metering for the types
• Check if GPU metrics can be supported already
  • Nvidia tutorial: https://blog.kubecost.com/blog/nvidia-gpu-usage/
  • OTel conventions: https://opentelemetry.io/docs/specs/semconv/system/hardware-metrics/
  • NVidia exporter: https://opentelemetry.io/docs/specs/semconv/system/hardware-metrics/
• Add support to KEB and document the new feature
• Enhance quota monitoring for the new types
• Adjust price calculator
• Do anything additional needed as result of the preparation (driver support?)