Add dranet blog

[anson627]

No description provided.

[Copilot]

Pull request overview

Adds a new AKS website blog post explaining how to use DRANET + Dynamic Resource Allocation (DRA) for NUMA-aware GPU/NIC alignment (RDMA performance) and includes accompanying control-plane/data-plane diagrams (Mermaid sources + exported SVGs).

Changes:

• New blog post: RDMA/NUMA scheduling problem statement, DRANET architecture, DRA ResourceClaimTemplate examples, and NCCL benchmark walkthrough/results.
• Added control-plane and data-plane diagrams as both .mmd sources and rendered .svg assets.

Reviewed changes

Copilot reviewed 3 out of 5 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
website/blog/2026-04-01-dranet-rdma-optimization-for-ai-on-aks/index.md	New blog post content + configuration examples + benchmark walkthrough
website/blog/2026-04-01-dranet-rdma-optimization-for-ai-on-aks/control-plane-diagram.mmd	Mermaid source for control-plane diagram
website/blog/2026-04-01-dranet-rdma-optimization-for-ai-on-aks/control-plane-diagram.svg	Rendered control-plane diagram used by the post
website/blog/2026-04-01-dranet-rdma-optimization-for-ai-on-aks/data-plane-diagram.mmd	Mermaid source for data-plane diagram
website/blog/2026-04-01-dranet-rdma-optimization-for-ai-on-aks/data-plane-diagram.svg	Rendered data-plane diagram used by the post

[Copilot]

Grammar: "In previous post" reads like a missing article. Consider changing to "In a previous post, we covered..." (or similar) for correct English.

Suggested change

In previous post, we covered [fundamental DRA concepts](/2025/11/17/dra-devices-and-drivers-on-kubernetes). In this post, we walk through how DRANET works on [AKS 1.34](https://kubernetes.io/blog/2025/09/01/kubernetes-v1-34-dra-updates/) with [ND GB300-v6](https://learn.microsoft.com/en-us/azure/virtual-machines/sizes/gpu-accelerated/nd-gb300-v6-series?tabs=sizebasic) nodes, demonstrate three NUMA (Non-uniform memory access) alignment scenarios, and show the benchmark results.

In a previous post, we covered [fundamental DRA concepts](/2025/11/17/dra-devices-and-drivers-on-kubernetes). In this post, we walk through how DRANET works on [AKS 1.34](https://kubernetes.io/blog/2025/09/01/kubernetes-v1-34-dra-updates/) with [ND GB300-v6](https://learn.microsoft.com/en-us/azure/virtual-machines/sizes/gpu-accelerated/nd-gb300-v6-series?tabs=sizebasic) nodes, demonstrate three NUMA (Non-uniform memory access) alignment scenarios, and show the benchmark results.

[Copilot]

The Microsoft Learn URL uses a locale-specific path (/en-us/). Repo blog guidance recommends using locale-agnostic Learn links (no /en-us/) to avoid unnecessary redirects and keep links consistent.

Suggested change

In previous post, we covered [fundamental DRA concepts](/2025/11/17/dra-devices-and-drivers-on-kubernetes). In this post, we walk through how DRANET works on [AKS 1.34](https://kubernetes.io/blog/2025/09/01/kubernetes-v1-34-dra-updates/) with [ND GB300-v6](https://learn.microsoft.com/en-us/azure/virtual-machines/sizes/gpu-accelerated/nd-gb300-v6-series?tabs=sizebasic) nodes, demonstrate three NUMA (Non-uniform memory access) alignment scenarios, and show the benchmark results.

In previous post, we covered [fundamental DRA concepts](/2025/11/17/dra-devices-and-drivers-on-kubernetes). In this post, we walk through how DRANET works on [AKS 1.34](https://kubernetes.io/blog/2025/09/01/kubernetes-v1-34-dra-updates/) with [ND GB300-v6](https://learn.microsoft.com/azure/virtual-machines/sizes/gpu-accelerated/nd-gb300-v6-series?tabs=sizebasic) nodes, demonstrate three NUMA (Non-uniform memory access) alignment scenarios, and show the benchmark results.

[Copilot]

Several Markdown tables start with || (double leading pipe), which renders as an extra empty first column in CommonMark/Docusaurus tables. Removing the extra leading | (use | ... | ... |) will make the tables render as intended.

[Copilot]

Pull request overview

Copilot reviewed 4 out of 7 changed files in this pull request and generated 4 comments.

[Copilot]

Same inconsistency here as in the earlier templates: deviceClassName: dranet.net should match the driver/DeviceClass identifier used elsewhere in the post (e.g., the dra.net driver/attribute namespace shown above).

[Copilot]

Pull request overview

Copilot reviewed 5 out of 8 changed files in this pull request and generated 1 comment.

[Copilot]

Pull request overview

Copilot reviewed 5 out of 8 changed files in this pull request and generated 2 comments.

[Copilot]

The ResourceSlice example defines attribute keys like dra.net/numaNode / dra.net/rdmaDevice with typed values (int, string, bool), but the CEL selectors later access them as device.attributes["dra.net"]["numaNode"] / ... ["rdmaDevice"] and compare directly to primitives. Please make the attribute schema and the selector syntax consistent (either update the ResourceSlice example to match the selector structure, or update selectors to reference the exact attribute keys/types shown in the ResourceSlice example).

Suggested change

	dra.net/numaNode:
	int: 0
	dra.net/pciAddress:
	string: "0101:00:00.0"
	dra.net/rdma:
	bool: true
	dra.net/rdmaDevice:
	string: mlx5_0
	dra.net/pciVendor:
	string: Mellanox Technologies
	dra.net:
	numaNode:
	int: 0
	pciAddress:
	string: "0101:00:00.0"
	rdma:
	bool: true
	rdmaDevice:
	string: mlx5_0
	pciVendor:
	string: Mellanox Technologies

[Copilot]

The CEL selector examples appear to assume attributes are nested under device.attributes["dra.net"] and directly comparable (e.g., == 0 / == true). If the published attributes follow the dra.net/<key>: {int|bool|string: ...} pattern shown earlier, these selectors won’t match as written. Please update the selector examples to the correct attribute access pattern so readers can copy/paste them successfully.

[Copilot]

Pull request overview

Copilot reviewed 5 out of 8 changed files in this pull request and generated 3 comments.

[Copilot]

The hardware-topology table is written with a leading || on each row (for example || Resource | Count | Detail |), which renders as an extra empty first column in Markdown. Convert these rows to standard table syntax (single leading |, or | | only when you intentionally need a blank header cell) so the table renders as intended.

[Copilot]

In the ResourceSlice example, attributes are shown with flat keys like dra.net/numaNode and dra.net/pciAddress, but later the CEL selectors access attributes as a nested map (device.attributes["dra.net"]["..."]). These formats are inconsistent; update the examples so the published attributes and selector expressions use the same schema.

Suggested change

	dra.net/numaNode:
	int: 0
	dra.net/pciAddress:
	string: "0101:00:00.0"
	dra.net/rdma:
	bool: true
	dra.net/rdmaDevice:
	string: mlx5_0
	dra.net/pciVendor:
	string: Mellanox Technologies
	dra.net:
	numaNode:
	int: 0
	pciAddress:
	string: "0101:00:00.0"
	rdma:
	bool: true
	rdmaDevice:
	string: mlx5_0
	pciVendor:
	string: Mellanox Technologies

[Copilot]

This section says the NIC devices come from the dranet.net DeviceClass, but elsewhere in the post the NIC driver/namespace is dra.net (for example driver: dra.net and dra.net/* attributes). Please clarify the intended identifiers (DeviceClass vs driver name) and make the examples consistent so readers can copy/paste them reliably.

Suggested change

With both drivers publishing ResourceSlices, we can write ResourceClaimTemplates that use CEL selectors to express precise GPU-NIC co-location constraints. Each template creates a per-pod ResourceClaim that requests devices from both the `gpu.nvidia.com` and `dranet.net` DeviceClasses, filtered by attributes like NUMA node or PCI address. We define three templates to demonstrate different NUMA placement strategies.

With both drivers publishing ResourceSlices, we can write ResourceClaimTemplates that use CEL selectors to express precise GPU-NIC co-location constraints. Each template creates a per-pod ResourceClaim that requests devices from both the `gpu.nvidia.com` DeviceClass and the DRANET NIC DeviceClass, `dranet.net`, filtered by attributes like NUMA node or PCI address. In these examples, `dranet.net` is the DeviceClass name, while `dra.net` is the DRANET driver name and attribute namespace used in the published device attributes and CEL selectors. We define three templates to demonstrate different NUMA placement strategies.

[Copilot]

Pull request overview

Copilot reviewed 5 out of 8 changed files in this pull request and generated 4 comments.

[Copilot]

The Microsoft Learn URL uses a locale-specific path (/en-us/). Repo guidance prefers locale-agnostic Learn links (for example, https://learn.microsoft.com/azure/...) to avoid unnecessary redirects and keep links consistent.

Suggested change

	In this post, we walk through how DRANET works on [AKS 1.34](https://kubernetes.io/blog/2025/09/01/kubernetes-v1-34-dra-updates/) with [Azure ND GB300-v6](https://learn.microsoft.com/en-us/azure/virtual-machines/sizes/gpu-accelerated/nd-gb300-v6-series?tabs=sizebasic) nodes, demonstrate three NUMA (Non-Uniform Memory Access) alignment scenarios, show and compare the RDMA benchmark results.
	In this post, we walk through how DRANET works on [AKS 1.34](https://kubernetes.io/blog/2025/09/01/kubernetes-v1-34-dra-updates/) with [Azure ND GB300-v6](https://learn.microsoft.com/azure/virtual-machines/sizes/gpu-accelerated/nd-gb300-v6-series?tabs=sizebasic) nodes, demonstrate three NUMA (Non-Uniform Memory Access) alignment scenarios, show and compare the RDMA benchmark results.

[Copilot]

Branding/capitalization is inconsistent: the paragraph uses "Nvidia ConnectX-5" but later the table uses "Mellanox ConnectX". Consider standardizing to the current vendor name (for example, "NVIDIA ConnectX-5" / "NVIDIA (Mellanox) ConnectX-5") for clarity and consistency.

Suggested change

	On an Azure ND GB300-v6 node, there are four NVIDIA GB300 GPUs and four [Nvidia ConnectX-5](https://www.nvidia.com/en-sg/networking/ethernet/connectx-5/) NICs with [InfiniBand](https://www.nvidia.com/en-us/networking/products/infiniband/) spread across two NUMA domains. The hardware topology looks like this:
	| Resource | Count | Detail |
	|---|---|---|
	| GPU | 4x NVIDIA GB300 | 288 GB HBM3E each |
	| NIC | 4x Mellanox ConnectX | 100 GB/s InfiniBand each |
	On an Azure ND GB300-v6 node, there are four NVIDIA GB300 GPUs and four [NVIDIA ConnectX-5](https://www.nvidia.com/en-sg/networking/ethernet/connectx-5/) NICs with [InfiniBand](https://www.nvidia.com/en-us/networking/products/infiniband/) spread across two NUMA domains. The hardware topology looks like this:
	| Resource | Count | Detail |
	|---|---|---|
	| GPU | 4x NVIDIA GB300 | 288 GB HBM3E each |
	| NIC | 4x NVIDIA ConnectX-5 | 100 GB/s InfiniBand each |

[Copilot]

This sentence has a double space ("wrong NIC") and awkward phrasing/punctuation (", can"). Consider revising to "the wrong NIC" and removing the extra space/comma to improve readability.

Suggested change

	Without topology-aware scheduling, Kubernetes has no way to co-locate a GPU and its NUMA-local NICs in the same ResourceClaim. Scheduling a workload onto a GPU with a wrong NIC on a different NUMA node, can silently result in slower data paths and degrade RDMA performance.
	Without topology-aware scheduling, Kubernetes has no way to co-locate a GPU and its NUMA-local NICs in the same ResourceClaim. Scheduling a workload onto a GPU with the wrong NIC on a different NUMA node can silently result in slower data paths and degrade RDMA performance.

[Copilot]

Section heading reads like a question but is missing the question mark. Either add "?" or change the heading to a statement like "How DRANET works".

Suggested change

	## How does DRANET work
	## How does DRANET work?