Skip to content
Master Your Finances — Finance Now's Market Pulse — Energy — Master Your Financial Future

Kubernetes: Get the Most from Dynamic Resource Allocation

DRA allows the scheduler to understand specific device attributes, setting the stage for "locality-aware" scheduling to minimize data latency.

,   and  Administrator
1 min read
In this image we can see the information board, buildings, shed, trees, electric cables and sky...
In this image we can see the information board, buildings, shed, trees, electric cables and sky with clouds.

Kubernetes: Get the Most from Dynamic Resource Allocation

Kubernetes 1.35 has strengthened its support for complex hardware workloads with a key update. The latest release includes the core components of Device Resource Allocation (DRA), now ready for production use. This feature aims to help organisations manage rising data centre costs while improving efficiency for AI and machine learning tasks.

DRA first graduated to stable status in Kubernetes 1.34 and is now always enabled in version 1.35. Developed by a Cloud Native Computing Foundation working group, it enhances the scheduler’s ability to assign jobs to specific hardware like GPUs, CPUs, network cards, and AI accelerators. Unlike older Device Plugins, DRA fills critical gaps by supporting improved scoring, device reuse in init containers, and binding conditions for smarter scheduling.

The feature also exposes device locality to the scheduler, allowing it to place workloads closer to the hardware they need. This locality-aware scheduling helps reduce latency and boosts performance for tasks like large language model (LLM) training, inference, and network-heavy applications. Users can now define exact hardware requirements—such as a specific GPU—and DRA will match the request to available devices. Future plans for DRA include expanded controls, such as managing hardware topologies. This will give administrators even finer control over resource allocation, ensuring high-priority jobs run on premium nodes while cost-sensitive tasks use cheaper spot instances. The goal is to maximise hardware use and reliability across mixed workloads.

With data centre electricity and hardware costs climbing, DRA offers a way to extract more value from existing infrastructure. The feature’s production-ready status in Kubernetes 1.35 provides organisations with a tool to optimise AI and ML workloads. By aligning jobs with the most suitable hardware, it aims to cut waste and improve efficiency in large-scale deployments.

Read also:

Related

Latest