With at least a concept of the structure of your application, it's difficult to see what alternatives there are to your problem, other than the one which you can't have. Maybe that's the point, but I was party to a solution to a similar problem of max'd out memory forcing a partially disc-based dataset which in resulted in an unacceptable reduction in performance.

The solution we arrived at was to partition both the dataset and the algorithms and run it in two halves on two seperate machines. Communication was done through a dedicated 100Mb/s FDDI link. There was talk of moving to a 155Mb/s ATM link, but this proved unnecessary.

It was initially thought that putting half the data and half the code on each box was the way forward, but in the end, the codeset was almost identical and actually gained some weight due to the (surprisingly small) extra to handle the communications.

The simple act of partitioning the dataset in two freed enough memory that the additional code was handled without penalty.

Just a possibility.