Add dranet blog

website/blog/2026-04-01-dranet-rdma-optimization-for-ai-on-aks/benchmark-chart.mmd

@@ -0,0 +1,11 @@
+---
+config:
+  theme: base
+  themeVariables:
+    primaryColor: "#9f62eb"
+---
+xychart-beta
+    title "NCCL all_reduce_perf — Avg Bus Bandwidth (GB/s)"
+    x-axis ["1nic-unaligned (cross-NUMA)", "1nic-aligned (same NUMA)", "2nic-aligned (same NUMA)"]
+    y-axis "Avg busbw (GB/s)" 0 --> 120
+    bar [25, 56, 112]

website/blog/2026-04-01-dranet-rdma-optimization-for-ai-on-aks/control-plane-diagram.mmd

@@ -0,0 +1,36 @@
+flowchart TB
+    subgraph User["Workload Author"]
+        RCT["ResourceClaimTemplate<br/>(CEL selectors)"]
+        PodSpec["Pod Spec<br/>(resourceClaims reference)"]
+    end
+
+    subgraph CP["Kubernetes Control Plane"]
+        API["API Server<br/>(DRA API group)"]
+        Sched["Scheduler<br/>(Topology-aware)"]
+        RS_GPU["ResourceSlice<br/>(gpu.nvidia.com)<br/>pciBusID, NUMA, pcieRoot"]
+        RS_NIC["ResourceSlice<br/>(dra.net)<br/>rdmaDevice, NUMA, pciAddress"]
+    end
+
+    subgraph Node["Kubernetes Node (Azure ND GB300-v6)"]
+        NVDRV["NVIDIA GPU DRA Driver<br/>(DaemonSet)"]
+        DRANETDRV["DRANET DRA Driver<br/>(DaemonSet)"]
+    end
+
+    %% User submits workload
+    PodSpec -->|"Submit pod with<br/>resource claims"| API
+    RCT -->|"Define GPU+NIC<br/>alignment constraints"| API
+
+    %% Drivers publish device topology
+    NVDRV -->|"Discover GPUs &<br/>publish topology"| RS_GPU
+    DRANETDRV -->|"Discover NICs &<br/>publish topology"| RS_NIC
+
+    %% Scheduler uses slices to allocate
+    RS_GPU --> Sched
+    RS_NIC --> Sched
+    Sched -->|"Evaluate CEL selectors"| API
+    API -->|"Bind pod to node<br/>with allocated devices"| Node
+
+    %% Styling
+    style User fill:#fef7e0,stroke:#fbbc04
+    style CP fill:#e8f0fe,stroke:#4285f4
+    style Node fill:#f3e8fd,stroke:#9f62eb

website/blog/2026-04-01-dranet-rdma-optimization-for-ai-on-aks/data-plane-diagram.mmd

@@ -0,0 +1,44 @@
+flowchart TB
+    Kubelet["Kubelet"]
+    CRI["containerd"]
+    NRI["NRI Plugin<br/>(DRANET)"]
+
+    subgraph NUMA0["NUMA Node 0"]
+        GPU0["GPU 0<br/>NVIDIA GB300"]
+        GPU1["GPU 1<br/>NVIDIA GB300"]
+        NIC0["NIC 0<br/>NVIDIA ConnectX-5"]
+        NIC1["NIC 1<br/>NVIDIA ConnectX-5"]
+    end
+
+    subgraph NUMA1["NUMA Node 1"]
+        GPU2["GPU 2<br/>NVIDIA GB300"]
+        GPU3["GPU 3<br/>NVIDIA GB300"]
+        NIC2["NIC 2<br/>NVIDIA ConnectX-5"]
+        NIC3["NIC 3<br/>NVIDIA ConnectX-5"]
+    end
+
+    subgraph Pod["Scheduled Pod"]
+        Container["Container<br/>/dev/infiniband/uverbs*"]
+    end
+
+    %% Runtime flow
+    Kubelet -->|"1. Receive device allocation<br/>result from API Server"| CRI
+    CRI -->|"2. Execute OCI CreateContainer<br/>hook"| NRI
+    NRI -->|"3. Inject allocated<br/>/dev/infiniband/* devices"| Pod
+
+    %% NUMA-aligned GDR paths
+    GPU0 <-.->|"PCIe · GDR ✓"| NIC0
+    GPU1 <-.->|"PCIe · GDR ✓"| NIC1
+    GPU2 <-.->|"PCIe · GDR ✓"| NIC2
+    GPU3 <-.->|"PCIe · GDR ✓"| NIC3
+
+    %% Cross-NUMA penalty
+    GPU0 <-.->|"QPI/UPI · No GDR ✗"| NIC3
+
+    %% Pod uses aligned devices
+    Container -.->|"4. NCCL uses<br/>GPU * + mlx5_*"| GPU0
+
+    %% Styling
+    style NUMA0 fill:#e6f4ea,stroke:#34a853
+    style NUMA1 fill:#fce8e6,stroke:#ea4335
+    style Pod fill:#fef7e0,stroke:#fbbc04