## Recap

You can use **Gunicorn** (or also Uvicorn) as a process manager with Uvicorn workers to take advantage of **multi-core CPUs**, to run **multiple processes in parallel**.

You could use these tools and ideas if you are setting up **your own deployment system** while taking care of the other deployment concepts yourself.

On the other hand, if you have 2 servers and you are using **100% of their CPU and RAM**, at some point one process will ask for more memory, and the server will have to use the disk as "memory" (which can be thousands of times slower), or even **crash**. Or one process might need to do some computation and would have to wait until the CPU is free again.

void TFE_SetLogicalCpuDevices(TFE_Context* ctx, int num_cpus,
                              const char* prefix, TF_Status* status) {
  std::vector<std::unique_ptr<tensorflow::Device>> devices;

  if (prefix == nullptr || strlen(prefix) == 0)
    prefix = "/job:localhost/replica:0/task:0";

  tensorflow::SessionOptions sess_options;
  (*sess_options.config.mutable_device_count())["CPU"] = num_cpus;
  status->status =

Wie viele Systemressourcen möchten Sie verbrauchen/nutzen? Sie mögen „nicht viel“ denken, aber in Wirklichkeit möchten Sie tatsächlich **so viel wie möglich ohne Absturz** verwenden.

		if nErr.Err != nil {
			peersLogOnceIf(logger.SetReqInfo(ctx, reqInfo), nErr.Err, nErr.Host.String())
		}
	}
}

// GetCPUs - Get all CPU information.
func (sys *NotificationSys) GetCPUs(ctx context.Context) []madmin.CPUs {
	reply := make([]madmin.CPUs, len(sys.peerClients))

	g := errgroup.WithNErrs(len(sys.peerClients))
	for index, client := range sys.peerClients {
		if client == nil {
			continue
		}

	return resp.ValueOrZero(), err
}

// GetCPUs - fetch CPU information for a remote node.
func (client *peerRESTClient) GetCPUs(ctx context.Context) (info madmin.CPUs, err error) {
	resp, err := getCPUsHandler.Call(ctx, client.gridConn(), grid.NewMSS())
	return resp.ValueOrZero(), err
}

  // Create a parallel device with two CPUs
  const char* first_device_name =
      "/job:localhost/replica:0/task:0/device:CUSTOM:0";
  std::array<const char*, 2> first_underlying_devices{
      "/job:localhost/replica:0/task:0/device:CPU:0",
      "/job:localhost/replica:0/task:0/device:CPU:1"};
  RegisterParallelDevice(context.get(), first_device_name,

- When the kubelet fails to assign CPUs to a Pod because there less available CPUs than the Pod requests, the error message changed from

// through TFE_ContextOptions, this method should not be called.
TF_CAPI_EXPORT extern void TFE_SetLogicalCpuDevices(TFE_Context* ctx,
                                                    int num_cpus,
                                                    const char* prefix,
                                                    TF_Status* status);

// Set configuration key and value using coordination service.

</p>

<p>
Note that the prohibition on introducing data races
does not apply if the compiler can prove that the races
do not affect correct execution on the target platform.
For example, on essentially all CPUs, it is valid to rewrite
</p>

<pre>
n := 0
for i := 0; i < m; i++ {
	n += *shared
}
</pre>

into:

<pre>
n := 0
local := *shared
for i := 0; i < m; i++ {
	n += local