This is a reference for understanding when a machine or instance is started up for your run, assuming your run is in a VPC deployment. To understand what the dispatcher is and how it works, please first review Cluster Controls. That documentation will describe in detail the Run Capacity, Max number of instances and Minimum number of instances seen on the dispatcher page.
Now that you understand how dispatcher works, let's walk through some user scenarios:
Let's say you have a hardware tier that has a 2 run max capacity per instance and you have a max instance count of 4. You currently have 1 instance spun up.
10 separate runs are launched simultaneously. In this scenario, the first two runs would be put on the 1 instance that is already available. Domino would then spin up 3 additional instances (stopping at the set max instance count of 4) and place 2 runs on each instance. Thus, instance 2 would get runs 3 and 4, instance 3 would get runs 5 and 6, and instance 4 would get runs 7 and 8. The runs 9 and 10 would be queued, waiting for a run slot to open on one of the existing instances. If you'd like to explore the different run states further you can see Why hasn't my run started yet. Also, in this scenario all 4 instances would accumulate one hour each of AWS billing time (since AWS charges by the hour).
For a second example, let's continue with the hardware tier we used in example one.
One run is launched and completes. Once this run is complete, a second run is launched and completes. Then, a third run is launched and completes. In this scenario, the first run will be put on the 1 instance that is already available. The second run will also be put on the 1st instance and the 3rd run will be placed on the 1st instance, too. In this example, no additional instances will be spun up because the single instance is sufficient for all 3 runs. At no time is there more than 1 run happening simultaneously, thus no additional instances need to be spun up.
Note: the configurable limit for the max number of instances that can be launched at once, is a security measure to keep someone from launching thousands of runs at once.