## 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。

// ContainerResourceMetricSource indicates how to scale on a resource metric known to
// Kubernetes, as specified in requests and limits, describing each pod in the
// current scale target (e.g. CPU or memory).  The values will be averaged
// together before being compared to the target.  Such metrics are built in to
// Kubernetes, and have special scaling options on top of those available to

  optional int32 desiredReplicas = 4;

  // currentCPUUtilizationPercentage is the current average CPU utilization over all pods, represented as a percentage of requested CPU,
  // e.g. 70 means that an average pod is using now 70% of its requested CPU.
  // +optional
  optional int32 currentCPUUtilizationPercentage = 5;
}

// MetricSpec specifies how to scale based on a single metric

      panels.timeSeries.bytes('Memory Usage', queries.memUsage, 'Memory usage of each running instance'),
      panels.timeSeries.base('CPU Usage', queries.cpuUsage, 'CPU usage of each running instance'),
    ]),
    row.new('Network')
    + row.withPanels([
      panels.timeSeries.connections('Connections', queries.connections, 'Connections opened and closed per instance'),

      panels.timeSeries.allocations('Memory Allocations', queries.goAllocations, 'Details about memory allocations'),
      panels.timeSeries.base('CPU Usage', queries.cpuUsage, 'CPU usage of each running instance'),
      panels.timeSeries.base('Goroutines', queries.goroutines, 'Goroutine count for each running instance'),
    ]),
  ], panelHeight=10, startY=1)
  + g.util.grid.makeGrid([

[main]
summary=Maximum server performance for MinIO

[vm]
transparent_hugepage=madvise

[sysfs]
/sys/kernel/mm/transparent_hugepage/defrag=defer+madvise
/sys/kernel/mm/transparent_hugepage/khugepaged/max_ptes_none=0

[cpu]
force_latency=1
governor=performance
energy_perf_bias=performance
min_perf_pct=100

[sysctl]
fs.xfs.xfssyncd_centisecs=72000
net.core.busy_read=50
net.core.busy_poll=50
kernel.numa_balancing=1

TFCI_BAZEL_COMMON_ARGS="--repo_env=HERMETIC_PYTHON_VERSION=$TFCI_PYTHON_VERSION --config release_arm64_linux"
TFCI_BAZEL_TARGET_SELECTING_CONFIG_PREFIX=linux_arm64
# Note: this is not set to "--cpu", because that changes the package name
# to tensorflow_cpu. These ARM builds are supposed to have the name "tensorflow"
# despite lacking Nvidia CUDA support.
TFCI_BUILD_PIP_PACKAGE_ARGS="--repo_env=WHEEL_NAME=tensorflow"

Nesse caso, seria melhor ter apenas 2 servidores e usar uma porcentagem maior de seus recursos (CPU, memória, disco, largura de banda de rede, etc).

In diesem Fall könnte es besser sein, nur zwei Server zu haben und einen höheren Prozentsatz von deren Ressourcen zu nutzen (CPU, Arbeitsspeicher, Festplatte, Netzwerkbandbreite, usw.).

if [[ "$TFCI_WHL_NUMPY_VERSION" == 1 ]]; then
  # Uninstall tf nightly wheel built with numpy1.
  "$python" -m pip uninstall -y tf_nightly_numpy1
  # Install tf nightly cpu wheel built with numpy2.x from PyPI in numpy1.x env.
  "$python" -m pip install tf-nightly-cpu
  if [[ "$TFCI_WHL_IMPORT_TEST_ENABLE" == "1" ]]; then
    "$python" -c 'import tensorflow as tf; t1=tf.constant([1,2,3,4]); t2=tf.constant([5,6,7,8]); print(tf.add(t1,t2).shape)'