DRANET is a Kubernetes Network Driver that uses Dynamic Resource Allocation (DRA) to deliver high-performance networking for demanding applications in Kubernetes.
- DRA Integration: Leverages the power of Kubernetes' Dynamic Resource Allocation.
- High-Performance Networking: Designed for demanding workloads like AI/ML applications.
- Simplified Management: Easy to deploy and manage.
- Enhanced Efficiency: Optimizes resource utilization for improved overall performance.
- Cluster-Wide Scalability: Effectively manages network resources across a large number of nodes for seamless operation in Kubernetes deployments.
Our research paper, "The Kubernetes Network Driver Model: A Composable Architecture for High-Performance Networking," provides a deep dive into the DraNet model and its impact.
The key findings include:
- Up to 60% Bandwidth Increase: By enabling topology-aware scheduling of GPUs and NICs, DraNet boosts bus bandwidth by up to 59.6% for
all_gatherand 58.1% forall_reduceoperations in distributed AI/ML workloads. - Operational Simplicity: The paper demonstrates how the KND model used by DraNet drastically simplifies the management of high-performance hardware, replacing fragile, multi-component chains with a clean, composable architecture.
Read the full research paper here!
The DraNet driver communicates with the Kubelet through the DRA API and with the Container Runtime via NRI. This architectural approach ensures robust supportability and minimizes complexity, making it fully compatible with existing CNI plugins in your cluster.
Upon the creation of a Pod's network namespaces, the Container Runtime initiates a GRPC call to DraNet via NRI to execute the necessary network configurations.
A more detailed diagram illustrating this process can be found in our documentation: How It Works.
To get started with DraNet, your Kubernetes cluster needs to have Dynamic Resource Allocation (DRA) enabled. DRA is beta and is disabled by default in Kubernetes v1.32. You will need to enable both the feature gates and the API groups for DRA until it reaches GA.
If you are using KIND, you can create a cluster with the following configuration:
kind: Cluster
apiVersion: kind.x-k8s.io/v1alpha4
nodes:
- role: control-plane
image: kindest/node:v1.33.1
- role: worker
image: kindest/node:v1.33.1
- role: worker
image: kindest/node:v1.33.1
featureGates:
# Enable the corresponding DRA feature gates
DynamicResourceAllocation: true
DRAResourceClaimDeviceStatus: true
runtimeConfig:
api/beta : trueThen to create the cluster:
kind create cluster --config kind.yamlFor instructions on setting up DRA on GKE, refer to the official documentation: Set up Dynamic Resource Allocation
Install the latest sta 9FDA ble version of DraNet using the provided manifest:
kubectl apply -f https://raw.githubusercontent.com/google/dranet/refs/heads/main/install.yamlOnce DraNet is running, you can inspect the network interfaces and their
attributes published by the drivers. Users can then create DeviceClasses,
ResourceClaims, and/or ResourceClaimTemplates to schedule pods and allocate
network devices.
For examples of how to use DraNet with DeviceClass and ResourceClaim to
attach network interfaces to pods, please refer to the Quick Start
guide.
We welcome your contributions! Please review our Contributor License Agreement and Google's Open Source Community Guidelines before you begin. All submissions require review via GitHub pull requests.
For detailed development instructions, including local development with KIND and troubleshooting tips, see our Developer Guide.
Explore more concepts and advanced topics:
- Design: Understand the architectural choices behind DraNet: Design
- RDMA: Learn about RDMA components in Linux and their interplay: RDMA
- References: A list of relevant Kubernetes Enhancement Proposals (KEPs) and presentations: References
This is not an officially supported Google product. This project is not eligible for the Google Open Source Software Vulnerability Rewards Program.