I'm putting together a bunch of useful collections for high-performance scenarios.
The first is a port of the Google B-tree library to C#. It is written using unsafe code, and benchmarks within 10% or so of the original C++ version. As of now, it only supports keys and values that don't contain managed object references. This allows a very fast, memory-efficient solution.
I'll be adding more collections in the coming weeks.