The dBrain.cloud platform has implemented HAMi (Heterogeneous Accelerator Middleware) to manage GPU resources within Kubernetes clusters. This middleware acts as a scheduler layer between the GPU driver and workloads, intercepting API calls to enable dynamic resource allocation rather than relying on rigid profiles.
Unlike traditional methods like MIG, HAMi allows administrators to define GPU memory and compute power limits based on specific task requirements. This enables multiple processes to run in parallel on a single physical GPU with proper isolation, moving away from the inefficient 'one task per GPU' model.
The integration of HAMi into the dBrain Console allows users to manage GPU resources using standard Kubernetes manifests, such as requests and limits. The platform's scheduler automatically determines placement based on the actual state of the cluster, eliminating the need for manual configuration.
The dBrain Console now includes monitoring tools for GPU nodes and workloads, displaying consumption graphs alongside CPU and memory metrics. Users can allocate or edit GPU slots directly within the interface for Deployments, StatefulSets, or DaemonSets, improving infrastructure transparency.
