General information about ML-platform product
Selectel's ML platform is a prepared infrastructure for implementing ML development processes: training, deployment of ML models, and so on. The infrastructure consists of software and hardware components that are configured and prepared for operation.
When selecting ML platform components, all available cloud server configurations are used. Once the platform is connected, it can be extended with its own software components. Tested use:
- ClearML;
- Kubeflow — more about installing Kubeflow.
There are no additional restrictions on Selectel's management of the ML platform cluster.
Platform components
By default, the ML platform consists of:
- from hardware components:
- cloud platform — base for Managed Kubernetes with NVIDIA® GPUs (Tesla T4, A2, A30, A100, A2000, A5000, GTX 1080, RTX 2080 Ti);
- program components:
- Managed Kubernetes clusters with a preset;
- domain to access the Managed Kubernetes cluster;
- SSO Keycloak — authorization in internal services of the platform;
- Prom Stack — monitoring of platform components;
- Forecastle is the home page of the platform;
- object storage — A repository of datasets and experiment data;
- Container Registry — container image storage.
In Managed Kubernetes clusters:
- drivers installed;
- annotation of nodes has been performed;
- added the necessary GPU resources for computation;
- network is configured, including Traefik Kubernetes Ingress.
When the ClearML platform is installed in a cluster, it is directly managed through the SDK, which is installed in the user's own IDE. ClearML uses cluster nodes to run ML experiments. The ClearML architecture allows for a variety of component configurations:
- A single Managed Kubernetes cluster for all ML tasks;
- several clusters of Managed Kubernetes — each for its own task (Inference and Training);
- connecting a dedicated server as a computational node for ML experiments.
Connect the platform
To connect the platform, leave a request on the website. We will help you choose its configuration for your project and calculate the cost.
Cost
The cost of ML-platform is calculated after processing the request and selecting the configuration. It depends only on the platform components: Managed Kubernetes cluster, object storage, Container Registry.