I'm running a stress test tool that's simulating 10 concurrent requests for an ArcGIS Server 10.1 (Windows) Map Service. I have the Map Server set to use a maximum pool of 4 instances.
I'm not using caching - these are full screen map images based on random coordinates within a bounding area of interest.
Regardless of whether I'm using File Geodatabase, PostGIS or SQL Server as the source of the data - the maximum number of instances that gets used is 2.
FYI - the requests are coming from separate threads and are staggered by 200ms with an in-between request delay of 1000ms. So I would expect it to fire on all cylinders during the test.
Any ideas why it never uses the maximum number of instances?