The Feature Selection Iterator has been a long time coming for model builder and I use it regularly. Whilst it helps simplify models (we don't need to use Python for looping) I have noticed that it's performance can take a hit when there is a large selection.
I run models that crunch large datasets and I often find that ArcGIS simply runs out of memory. A solution to this is to select a subset of my data and crunch in batches. OK it means I have to keep an eye on everything and restart the processing but it gets the job done. If I pre-select a lot of data I notice the Iterator takes much longer to run it's make featurelayer routine. So as a real work example I have a File Geodatabase point layer with 12,000 points each needs to be processed individually. If I chuck it at my model with no selection it eventually runs out of memory. If I select say 500 points then the iterator's make feature layer step takes about 9 seconds before it moves onto the next model process. If I select say 2000 points (this would not crash ArcGIS) it takes about 45 seconds to make feature layer.
