HAMi 技术文档索引

flowchart LR subgraph user["用户层"] A["Pod Submit (kubectl apply)"] end subgraph controlplane["Kubernetes 控制平面"] B["API Server"] C["Mutating Webhook"] D["kube-scheduler"] end subgraph hamischeduler["HAMi Scheduler (Scheduler Extender)"] E["Filter (节点筛选)"] F["Bind (设备绑定)"] end subgraph nodecomponents["节点组件"] G["kubelet"] H["Device Plugin (ListAndWatch / Allocate)"] I["Container Runtime"] end subgraph container["容器内部"] J["libvgpu.so (LD_PRELOAD 拦截)"] K["用户应用"] end subgraph observability["可观测性"] L["vGPU Monitor"] M["Prometheus"] end A -->|"创建 Pod"| B B -->|"Admission Review"| C C -->|"注入 vGPU 资源请求"| B B -->|"调度请求"| D D -->|"Extender Filter"| E E -->|"返回可用 节点列表"| D D -->|"Extender Bind"| F F -->|"写入设备分配 Annotation"| B B -->|"绑定节点"| G G -->|"Allocate 请求"| H H -->|"配置环境变量 挂载 libvgpu.so"| I I --> J J -->|"API 拦截 显存/算力限制"| K L -->|"采集 vGPU 使用指标"| M style user fill:#e8f5e9,stroke:#4caf50,color:#000 style controlplane fill:#e3f2fd,stroke:#2196f3,color:#000 style hamischeduler fill:#fff3e0,stroke:#ff9800,color:#000 style nodecomponents fill:#f3e5f5,stroke:#9c27b0,color:#000 style container fill:#fce4ec,stroke:#e91e63,color:#000 style observability fill:#e0f2f1,stroke:#009688,color:#000

flowchart TB subgraph hami_repo["HAMi 仓库 (Go)"] direction TB S["Scheduler Extender pkg/scheduler/"] W["Mutating Webhook pkg/webhook/"] DP["Device Plugin pkg/device-plugin/"] DM["Device Manager pkg/device/"] MON["vGPU Monitor pkg/monitor/"] API["API Types pkg/api/"] UTIL["Util & Config pkg/util/"] end subgraph core_repo["HAMi-core 仓库 (C)"] direction TB HOOK["CUDA Hook Layer src/cuda/"] MEM["Memory Manager src/memory/"] COMPUTE["Compute Limiter src/compute/"] SHM["Shared Memory IPC src/shm/"] LIB["libvgpu.so (最终产物)"] end S -->|"Annotation 传递设备分配"| DP DP -->|"挂载 + 环境变量"| LIB MON -->|"读取共享内存"| SHM HOOK --> MEM HOOK --> COMPUTE MEM --> SHM COMPUTE --> SHM style hami_repo fill:#e8eaf6,stroke:#3f51b5,color:#000 style core_repo fill:#fbe9e7,stroke:#ff5722,color:#000

classDiagram class Devices { <> +CommonWord() string +MutateAdmission(ctr *v1.Container, pod *v1.Pod) bool +CheckHealth(devType string, n *v1.Node) (bool, bool) +NodeCleanUp(nn string) error +GetNodeDevices(n v1.Node) ([]*api.DeviceInfo, error) +CheckType(annos map[string]string, d util.DeviceUsage, n util.ContainerDeviceRequest) (bool, bool, bool) +CheckUUID(annos map[string]string, d util.DeviceUsage) bool +LockNode(n *v1.Node, pod *v1.Pod) error +ReleaseNodeLock(n *v1.Node, pod *v1.Pod) error +Allocate(devs []*api.DeviceInfo, reqs util.ContainerDeviceRequest) ([]*api.DeviceInfo, error) +GenerateResourceRequests(ctr *v1.Container) util.ContainerDeviceRequest end class DeviceUsage { +ID string +Index uint +Used int32 +Count int32 +Usedmem int32 +Totalmem int32 +Totalcore int32 +Usedcores int32 +Numa int +Type string +Health bool } class DeviceInfo { +ID string +Index uint +Count int32 +Devmem int32 +Devcore int32 +Type string +Numa int +Health bool } class ContainerDeviceRequest { +Nums int32 +Type string +Memreq int32 +MemPercentagereq int32 +Coresreq int32 } class ContainerDevice { +Idx int +UUID string +Type string +Usedmem int32 +Usedcores int32 } class PodDevices { +Name string +UID string +Devices map~string~ContainerDevice[][] } class NodeInfo { +ID string +Node *v1.Node +Devices DeviceUsageList +Score float32 } class NodeScore { +NodeID string +Devices DeviceUsageList +Score float32 } Devices ..> DeviceInfo : "返回设备信息" Devices ..> ContainerDeviceRequest : "生成资源请求" NodeInfo --> DeviceUsage : "包含多个设备" NodeScore --> DeviceUsage : "评分关联设备" PodDevices --> ContainerDevice : "容器设备映射" ContainerDeviceRequest ..> DeviceUsage : "匹配检查"