Admission control guarantees vSphere HA failover by ensuring enough spare failover capacity inside the cluster. You need free resources when hosts must manage the failure for all affected VMs and admission control takes care of you.

Configuration is at the cluster level under the vSphere Availability section shown as follows:

You can define host failover capacity using one of the following options:

• Disabled: This option is not a very good idea, but it is possible. For example, for a two nodes cluster, you may want to disable admission control during maintenance operation.
• Slot Policy (powered-on VMs): A slot is a logical representation of memory and CPU resources. A slot is the memory and CPU reservation required for any powered-on VMs in the cluster. Slot policy can do good work in the environment where there is a very similar VM with the same CPU memory configuration. When you have a lot of small VMs and two monster VMs, it is not a very good situation because the reservation is selecting the largest value. You can change this value for CPU slot size and memory slot size through the Advanced Options: 

• Cluster resource percentage: You can configure vSphere HA to perform admission control by reserving a specific percentage of cluster CPU and memory resources for recovery from host failures. vSphere HA calculates CPU and memory. CPU calculation uses CPU reservation for the powered-on VMs. If you don't use reservation HA, use default value 32 MHz. The memory calculates the memory reservation and memory overhead of each powered-on VM, the default value is 0 MB. You can override calculated failover capacity:

• Dedicated failover hosts: The last option is defined failover hosts. You can specify and dedicate failover hosts. vSphere HA uses these hosts when it needs failover actions or has insufficient resources:

Then, there is a new option, named Performance degradation VMs tolerate. This new setting in vSphere 6.5, if set, will issue a warning when a host failure would cause a reduction in VM performance based on the actual resource, not just configured reservations.