靈駿智算資源為您提供高性能AI訓練、高性能計算所需的異構計算算力服務，用于提交DLC訓練任務或創建DSW實例。本文為您介紹PAI靈駿智算服務從開通到使用的最佳實踐。

前提條件

• 已開通PAI（DLC、DSW）后付費，詳情請參見開通并創建默認工作空間。

• 已創建VPC、交換機和安全組，詳情請參見創建和管理專有網絡和創建安全組。

  說明

  如果您有訪問公網的需求，需要為該VPC開通公網訪問能力，具體操作，請參見使用公網NAT網關SNAT功能訪問互聯網。

• 已提交工單，申請添加靈駿智算使用白名單。

• 如果使用RAM角色提交DLC分布式訓練任務，需要完成信任策略配置，配置詳情請參見RAM角色登錄并使用PAI。

  其中為Role配置的權限策略在當前的實現中會影響到PAI DLC的任務運行，您可參考以下權限策略制定您的自定義策略（AliyunPAIDLCDefaultRolePolicy）：

  單擊此處，查看自定義權限策略

  {
      "Version": "1",
      "Statement": [
          {
              "Action": [
                  "cs:GetClusterById",
                  "cs:GetClusters",
                  "cs:GetUserConfig",
                  "cs:DescribeClusterNodes",
                  "cs:DescribeClusterInnerServiceKubeconfig",
                  "cs:RevokeClusterInnerServiceKubeconfig"
              ],
              "Resource": "*",
              "Effect": "Allow"
          },
          {
              "Action": [
                  "vpc:DescribeNatGateways",
                  "vpc:DescribeVpnGateways",
                  "vpc:DescribeEipAddresses",
                  "vpc:DescribeRouteTableList",
                  "vpc:DescribeIpv6Gateways",
                  "vpc:DescribeHaVips",
                  "vpc:DescribeVRouters",
                  "vpc:DescribeRouterInterfaceAttribute",
                  "vpc:DescribeSnatTableEntries",
                  "vpc:DescribeNetworkAcls",
                  "vpc:DescribeRegions",
                  "vpc:DescribeVpcs",
                  "vpc:DescribeVSwitches",
                  "vpc:DescribeVpcAttribute",
                  "vpc:DescribeVSwitchAttributes",
                  "vpc:AssociatePhysicalConnectionToVirtualBorderRouter",
                  "vpc:UnassociatePhysicalConnectionFromVirtualBorderRouter",
                  "vpc:DescribeAccessPoints",
                  "vpc:CreatePhysicalConnectionNew",
                  "vpc:DescribePhysicalConnections",
                  "vpc:EnablePhysicalConnection",
                  "vpc:CreateVirtualBorderRouter",
                  "vpc:CreateRouterInterface",
                  "vpc:ModifyRouterInterfaceAttribute",
                  "vpc:ModifyRouterInterfaceSpec",
                  "vpc:DescribeRouterInterfaces",
                  "vpc:DescribeRouteTables",
                  "vpc:ConnectRouterInterface",
                  "vpc:CreateRouteEntry",
                  "vpc:DeleteRouteEntry",
                  "vpc:DeactivateRouterInterface",
                  "vpc:DeleteRouterInterface",
                  "vpc:DeleteVirtualBorderRouter",
                  "vpc:TerminatePhysicalConnection",
                  "vpc:DeletePhysicalConnection",
                  "vpc:CreateVSwitch",
                  "vpc:DeleteVSwitch",
                  "vpc:ModifyVSwitchAttribute",
                  "vpc:DescribeVirtualBorderRouters"
              ],
              "Resource": "*",
              "Effect": "Allow"
          },
          {
              "Action": [
                  "ecs:DescribeSecurityGroupAttribute",
                  "ecs:DescribeSecurityGroups",
                  "ecs:CreateNetworkInterface",
                  "ecs:DeleteNetworkInterface",
                  "ecs:DescribeNetworkInterfaces",
                  "ecs:CreateNetworkInterfacePermission",
                  "ecs:DescribeNetworkInterfacePermissions",
                  "ecs:DeleteNetworkInterfacePermission"
              ],
              "Resource": "*",
              "Effect": "Allow"
          },
          {
              "Action": [
                  "nas:DescribeFileSystems",
                  "nas:CreateMountTarget",
                  "nas:DescribeMountTargets",
                  "nas:ModifyMountTarget",
                  "nas:DescribeProtocolMountTarget"
              ],
              "Resource": "*",
              "Effect": "Allow"
          },
          {
              "Action": [
                  "cr:ListNamespace",
                  "cr:ListRepository",
                  "cr:GetAuthorizationToken",
                  "cr:ListInstanceEndpoint",
                  "cr:PullRepository",
                  "cr:PushRepository",
                  "cr:GetInstance",
                  "cr:GetInstanceVpcEndpoint",
                  "cr:ListInstance",
                  "cr:ListInstanceDomain"
              ],
              "Resource": "*",
              "Effect": "Allow"
          },
          {
              "Action": [
                  "datasetacc:DescribeInstance",
                  "datasetacc:DescribeSlot",
                  "datasetacc:DescribeEndpoint"
              ],
              "Resource": "*",
              "Effect": "Allow"
          }
      ]
  }

使用限制

• 僅華北6（烏蘭察布）地域支持使用靈駿智算資源。

• 支持的任務類型

  當前靈駿智算資源僅支持TensorFlow、PyTorch、ElasticBatch和MPIJob類型的訓練任務。

• 當前僅智算CPFS類型的數據集暫不支持加速。