ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

  std::vector<std::string> devices{
      "/job:localhost/replica:0/task:0/device:CPU:0",
      "/job:localhost/replica:0/task:0/device:CPU:1"};
  ParallelDevice parallel_device(std::move(devices));
  std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> handle_op(
      TFE_NewOp(context.get(), "VarHandleOp", status.get()), TFE_DeleteOp);

  Args:
    environ_cp: copy of the os.environ.
    var_name: string for name of environment variable, e.g. "TF_NEED_CUDA".
    query_item: string for feature related to the variable, e.g. "CUDA for
      Nvidia GPUs".
    enabled_by_default: boolean for default behavior.
    question: optional string for how to ask for user input.
    yes_reply: optional string for reply when feature is enabled.

global:
  proxy:
    resources:
      requests:
        cpu: 10m
        memory: 40Mi

pilot:
  autoscaleEnabled: false
  traceSampling: 100
  resources:
    requests:
      cpu: 10m
      memory: 100Mi

gateways:
  istio-egressgateway:
    autoscaleEnabled: false
    resources:
      requests:
        cpu: 10m
        memory: 40Mi
  istio-ingressgateway:
    autoscaleEnabled: false

    kind: Deployment
    name: istiod{{- if not (eq .Values.revision "") }}-{{ .Values.revision }}{{- end }}
  metrics:
  - type: Resource
    resource:
      name: cpu
      target:
        type: Utilization
        averageUtilization: {{ .Values.pilot.cpu.targetAverageUtilization }}
  {{- if .Values.pilot.memory.targetAverageUtilization }}
  - type: Resource
    resource:
      name: memory
      target:

// Create a multi-client ServerDef with the given `job_name`, add `num_tasks`
// tasks and `num_virtual_gpus` virtual GPUs in it.
tensorflow::ServerDef GetMultiClientServerDef(const std::string& job_name,
                                              int num_tasks,
                                              int num_virtual_gpus = 0);

// Create a variable handle with name `variable_name` on a device with name
// `device_name`.

global:
  proxy:
    resources:
      requests:
        cpu: 10m
        memory: 40Mi

pilot:
  autoscaleEnabled: false
  traceSampling: 100
  resources:
    requests:
      cpu: 10m
      memory: 100Mi

gateways:
  istio-egressgateway:
    autoscaleEnabled: false
    resources:
      requests:
        cpu: 10m
        memory: 40Mi
  istio-ingressgateway:
    autoscaleEnabled: false

global:
  proxy:
    resources:
      requests:
        cpu: 10m
        memory: 40Mi

pilot:
  autoscaleEnabled: false
  traceSampling: 100
  resources:
    requests:
      cpu: 10m
      memory: 100Mi

gateways:
  istio-egressgateway:
    autoscaleEnabled: false
    resources:
      requests:
        cpu: 10m
        memory: 40Mi
  istio-ingressgateway:
    autoscaleEnabled: false

## 资源利用率

您的服务器是一个**资源**，您可以通过您的程序消耗或**利用**CPU 上的计算时间以及可用的 RAM 内存。

您想要消耗/利用多少系统资源？ 您可能很容易认为“不多”，但实际上，您可能希望在不崩溃的情况下**尽可能多地消耗**。

如果您支付了 3 台服务器的费用，但只使用了它们的一点点 RAM 和 CPU，那么您可能**浪费金钱** 💸，并且可能 **浪费服务器电力** 🌎，等等。

在这种情况下，最好只拥有 2 台服务器并使用更高比例的资源（CPU、内存、磁盘、网络带宽等）。

另一方面，如果您有 2 台服务器，并且正在使用 **100% 的 CPU 和 RAM**，则在某些时候，一个进程会要求更多内存，并且服务器将不得不使用磁盘作为“内存” （这可能会慢数千倍），甚至**崩溃**。 或者一个进程可能需要执行一些计算，并且必须等到 CPU 再次空闲。

  TFE_DeleteTensorHandle(retvals[0]);
  TFE_DeleteContext(ctx);
  TF_DeleteStatus(status);
}
TEST(CAPI, Execute_MatMul_CPU_Runtime_Error) {
  Execute_MatMul_CPU_Runtime_Error(false);
}
TEST(CAPI, Execute_MatMul_CPU_Runtime_ErrorAsync) {
  Execute_MatMul_CPU_Runtime_Error(true);
}

void Execute_MatMul_CPU_Type_Error(bool async) {
  TF_Status* status = TF_NewStatus();
  TFE_ContextOptions* opts = TFE_NewContextOptions();

  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  const char remote_device_name[] =
      "/job:localhost/replica:0/task:1/device:CPU:0";
  const char local_device_name[] =
      "/job:localhost/replica:0/task:0/device:CPU:0";
  CheckRemoteMatMulExecutesOK(ctx, remote_device_name, local_device_name);

  TFE_Executor* executor = TFE_ContextGetExecutorForThread(ctx);