Briefly - use storage local to processing units rather than remote storage.  Requires considerable refactoring to achieve -- EddieEdwards
-----
Involves considering a system architecture (such as the Playstation2) that values thoroughput over latency, bus bandwidth over cache size.  In an age of growing interest in the power of parallelization, it is an important consideration.  Note that it scales in usefulness from assembly optimization (use registers) all the way up to application layer design.  For instance, when writing a server for a MMORPG, it could be very important to utilize local world-state data to reduce the amount of network synchronization needed.  This is seen in Quake 3 where moving objects are 'predicted' (interpolated from local conditions/most recent synchs) in the case that network latency is high. -- KarlinFox