I got it wrong, let me tell you. After finishing the CEDB .NET code, I started to write the Pocket Access layer for .NET CF and realized how some of my early assumptions were simply wrong.
The CEDB code implements a record containing a collection of properties. Each property maps to a CEPROPVAL structure directly, meaning that the contained value is accessed through bit conversion methods. This is pretty stupid. My early reasoning was that this would make it so much easier to read and write the binary info to and from the database.
When I tried to fill a DataTable object using this approach, I was surprised with the poor speed. At first I blamed the DataGrid but realised that, although it is a slow loader, the bulk of the time was being spent loading the DataTable. I turned my attention to the DataTable and applied all the tricks of the trade to make it load faster. Nah... no way.
But wait! Could it be...? Was it in my code to load the records? Like Michael Abrash used to say: "Assume nothing: Test everything".
Bingo! As any Scorpio worth his salt, I rewrote the whole thing. Upon loading, the record now converts all its properties to boxed values or string references (blobs are just byte arrays, but references as well). These are all packed into an object array which makes it a breeze to load into a DataRow. Buffers are reused to avoid heap fragmentation and the undesirable visit of our friend the garbage collector.
The Suppliers table of the converted Northwind database now gets loaded into a brand-new DataTable object in under 200 ticks on my iPAQ 3850 (I cannot get enough of this machine...), down from almost 600 ticks using the old code. Impressive, huh?