How many desktops do you have on that host?
With 16 cores (I guess its 4x4cores) and desktops with 8 cores you will be over committing the host a lot if running more then 3-4 desktops.
When you did your tests was it with only one desktop or several?
I would suggest to lower the amount of vCPU:s to 4, its not likely that the application will be suffering since desktop applications are usually very bad at threading.
// Linjo