Automatic analytical modeling for the estimation of cache misses

Caches play a very important role in the performance of modern computer systems due to the gap between the memory and the processor speed. Among the methods for studying their behavior, the most widely used by now has been trace-driven simulation. Nevertheless, analytical modeling gives more information and requires smaller computation times that allow it to be used in the compilation step to drive automatic optimizations on the code. The traditional draw-back of analytical modeling has been its limited precision and the lack of techniques to apply it systematically without user intervention. In this work we present a methodology to build analytical models for codes with regular access patterns. These models can be applied to caches with an arbitrary size, line size and associativity. Their validation through simulations using typical scientific code fragments has proved a good degree of accuracy.