问题
When starting a TensorFlow server (tf.distribute.Server), you must pass a ClusterSpec which specifies all tasks/workers in the cluster.
Why? What does it use that information for?
Is this just that device names (for tf.device) can uniquely identify on what worker they run in the cluster? E.g. for with tf.device("/job:ps/task:0") or with tf.device("/job:worker/task:7")?
来源：https://stackoverflow.com/questions/62004631/why-does-a-tf-server-needs-to-know-about-all-other-tasks-workers-in-the-cluster