I have a legacy application that is written in C# and it displays a very complex treeview with 10 to 20 thousand elements.
In the past I encountered a similar problem (but in C++) that i solved with the OWNERDATA capability offered by the Win32 API.
Is there a similar mechanism in C#?
EDIT: The plan is to optimize the creation time as well as browsing time. The method available through Win32 API is excellent in both of these cases as it reduce initialization time to nothing and the number of requests for elements are limited to only the ones visible at any one time. Joshl: We are actually doing exactly what you suggest already, but we still need more efficiency.
I don't believe the .NET TreeView supports what you want, although this type of model is supported by .NET's DataGridView (see DataGridView's VirtualMode property). The TreeView will let you draw your own nodes but it won't let you populate them from some virtual store.
If possible, you might want to consider the use of a DataGridView for your application. If not, managing the nodes manually (like joshl mentions above) might work if you can get around some issues with refreshing the screen properly when nodes are expanded. Outside of that, you might want to check out some of the third party vendors, like this one (Divelements SandGrid), that might (emphasis on might) support your desired mode of operation.
NOTE: The SandGrid is not supported by Divelements as of the end of July 2013.