cnvrg.io has created an architecture with an ecosystem that enables an organization to build and deploy ML and AI on many infrastructures, including Dell optimized infrastructure and Dell Validated Designs for AI, as shown in the following figure:
Figure 3. The cnvrg.io architecture (source: cnvrg.io)
This ecosystem is enabled by the cnvrg.io Operating System for AI that consists of the following components:
- Control Plane—The Control Plane is the management layer designed to operate with any CNCF-compliant Kubernetes distribution. The control plane provides a “single pane of glass” to manage all the elements in the ML stack, including datasets, model code, jobs, model performance, cluster, and resource statistics. From the control plane, a data scientist has a consolidated view of all the elements open for manipulation in the project.
- AI Library—The AI Library is a package manager for the algorithms and data components. cnvrg.io integrates with GitHub, allowing data scientists to add their own custom repositories as well.
- Pipelines—The pipelines are where the all-important specialized work of the data scientist occurs. The drag-and-drop console enables a data scientist to build a complete end-to-end ML process that begins with data and ends with model serving and monitoring.
- Orchestration and Scheduling—Through the Kubernetes-based metascheduler, cnvrg.io enables all the tools for orchestrating pods, containers, jobs, and scaling resources across clusters.
- Compute and Storage—The compute and storage layers enable assignment of the underlying platform elements to pipeline stages, optimizing for the best type of compute, whether it be CPU, GPU, or any other specialized compute element.