HAMi 在 CNCF 生态中的定位

flowchart TB subgraph cncf["CNCF 项目层级"] direction TB graduated["Graduated (毕业)"] incubating["Incubating (孵化)"] sandbox["Sandbox (沙箱)"] end subgraph graduated_projects["毕业项目"] k8s["Kubernetes"] prometheus_g["Prometheus"] containerd_g["containerd"] end subgraph incubating_projects["孵化项目"] volcano_i["Volcano"] end subgraph sandbox_projects["沙箱项目"] hami["HAMi"] end graduated --> graduated_projects incubating --> incubating_projects sandbox --> sandbox_projects style hami fill:#4CAF50,color:#fff,stroke:#388E3C,stroke-width:2px style sandbox fill:#FFF9C4 style incubating fill:#BBDEFB style graduated fill:#C8E6C9

flowchart LR subgraph gap["HAMi 填补的技术缺口"] direction TB native["Kubernetes 原生能力"] native --> n1["CPU/Memory - 精细调度"] native --> n2["Storage - CSI 标准"] native --> n3["Network - CNI 标准"] native --> n4["GPU - 仅整卡分配"] n4 -->|"HAMi 增强"| enhanced["HAMi 补充能力"] enhanced --> e1["GPU 细粒度共享"] enhanced --> e2["显存/算力隔离"] enhanced --> e3["多厂商统一管理"] enhanced --> e4["可观测性集成"] end style n4 fill:#FFCDD2 style enhanced fill:#C8E6C9

flowchart TB subgraph k8s_scheduler["kube-scheduler 调度流程"] direction TB queue["调度队列"] prefilter["PreFilter"] filter_phase["Filter (过滤)"] score_phase["Score (评分)"] reserve["Reserve (预留)"] bind_phase["Bind (绑定)"] end subgraph extender_points["HAMi 扩展点"] ext_filter["Extender Filter POST /filter"] ext_bind["Extender Bind POST /bind"] end subgraph webhook_point["HAMi Webhook"] mut_webhook["Mutating Webhook POST /webhook"] end subgraph device_plugin_point["HAMi Device Plugin"] dp_api["Device Plugin API gRPC Allocate"] end queue --> prefilter prefilter --> filter_phase filter_phase -->|"HTTP 调用"| ext_filter ext_filter -->|"返回过滤结果"| filter_phase filter_phase --> score_phase score_phase --> reserve reserve --> bind_phase bind_phase -->|"HTTP 调用"| ext_bind ext_bind -->|"返回绑定结果"| bind_phase mut_webhook -.->|"Pod 创建时拦截 注入 schedulerName"| queue dp_api -.->|"Pod 启动时 注入 libvgpu.so"| bind_phase style ext_filter fill:#2196F3,color:#fff style ext_bind fill:#2196F3,color:#fff style mut_webhook fill:#FF9800,color:#fff style dp_api fill:#4CAF50,color:#fff

flowchart LR subgraph hami_monitor["HAMi vGPU Monitor"] metrics_endpoint[":9394/metrics Prometheus HTTP 端点"] shm_reader["共享内存读取器 读取 libvgpu.so 写入的数据"] end subgraph prometheus_stack["Prometheus 生态"] prometheus["Prometheus Server 指标采集与存储"] alertmanager["Alertmanager 告警管理"] grafana["Grafana 可视化面板"] servicemonitor["ServiceMonitor (Prometheus Operator)"] end subgraph dcgm_stack["NVIDIA 原生监控"] dcgm["DCGM Exporter 物理 GPU 指标"] end shm_reader --> metrics_endpoint prometheus -->|"Scrape"| metrics_endpoint prometheus -->|"Scrape"| dcgm prometheus --> alertmanager prometheus --> grafana servicemonitor -->|"自动发现"| metrics_endpoint style metrics_endpoint fill:#E65100,color:#fff

flowchart TB subgraph capability["能力矩阵"] direction LR subgraph hami_cap["HAMi 能力"] h1["GPU 细粒度共享"] h2["显存隔离"] h3["算力隔离"] h4["多厂商支持"] end subgraph volcano_cap["Volcano 能力"] v1["Gang Scheduling 组调度"] v2["Fair Scheduling 公平调度"] v3["Queue Management 队列管理"] v4["Job Lifecycle 作业生命周期"] end subgraph combined["组合能力"] c1["细粒度 GPU 共享 + 组调度"] c2["多租户 GPU 配额 + 公平调度"] c3["异构加速器管理 + 批处理优化"] end hami_cap --> combined volcano_cap --> combined end

flowchart TB subgraph app_layer["应用层"] ai_framework["AI 框架 PyTorch / TensorFlow / MindSpore"] inference_engine["推理引擎 vLLM / Triton / TensorRT"] end subgraph orchestration["编排与调度层"] k8s_core["Kubernetes (Graduated)"] volcano_orch["Volcano (Incubating) 批处理调度"] end subgraph device_mgmt["设备管理层"] hami_mgmt["HAMi (Sandbox) GPU 虚拟化与共享"] device_plugin_api["Device Plugin API (K8s 标准)"] dra_api["DRA API (K8s 1.30+ Alpha)"] end subgraph runtime_layer["运行时层"] containerd_rt["containerd (Graduated)"] nvidia_toolkit["nvidia-container-toolkit 容器 GPU 支持"] hamicore["HAMi-core libvgpu.so CUDA API Hook"] end subgraph observability_layer["可观测性层"] prometheus_obs["Prometheus (Graduated) 指标采集"] grafana_obs["Grafana 可视化"] otel["OpenTelemetry (Incubating) 分布式追踪"] end subgraph hardware["硬件层"] nvidia_hw["NVIDIA GPU A100/H100/V100"] ascend_hw["华为 Ascend NPU 910/310"] cambricon_hw["寒武纪 MLU 思元系列"] other_hw["其他加速器 AMD/海光/天数智芯..."] end ai_framework --> k8s_core inference_engine --> k8s_core k8s_core --> volcano_orch k8s_core --> hami_mgmt hami_mgmt --> device_plugin_api hami_mgmt --> dra_api device_plugin_api --> containerd_rt containerd_rt --> nvidia_toolkit nvidia_toolkit --> hamicore hamicore --> nvidia_hw hami_mgmt --> ascend_hw hami_mgmt --> cambricon_hw hami_mgmt --> other_hw hami_mgmt -->|"指标暴露"| prometheus_obs prometheus_obs --> grafana_obs hami_mgmt -.->|"规划中"| otel style hami_mgmt fill:#4CAF50,color:#fff,stroke:#388E3C,stroke-width:3px style hamicore fill:#F44336,color:#fff style k8s_core fill:#326CE5,color:#fff style volcano_orch fill:#1976D2,color:#fff style prometheus_obs fill:#E65100,color:#fff

flowchart LR subgraph stack["AI 基础设施技术栈"] direction TB l5["L5 - AI 应用层 LLM / CV / NLP / ASR"] l4["L4 - AI 框架层 PyTorch / TF / MindSpore / PaddlePaddle"] l3["L3 - 编排调度层 Kubernetes + Volcano + HAMi"] l2["L2 - 容器运行时层 containerd + HAMi-core (libvgpu.so)"] l1["L1 - 硬件驱动层 NVIDIA Driver / Ascend Driver / MLU Driver"] l0["L0 - 硬件层 GPU / NPU / MLU / DCU"] end l5 --> l4 l4 --> l3 l3 --> l2 l2 --> l1 l1 --> l0 style l3 fill:#4CAF50,color:#fff,stroke:#388E3C,stroke-width:2px style l2 fill:#F44336,color:#fff