Quantifying the Influences of Data Prefetching Using Artificial Neural Networks
Kecheng Ji, Li Liu · 2018
Data prefetching has been widely used in modern cache subsystems.Actually, an aggressive prefetching may bring negative yields unexpectedly, in which a new proposed prefetching strategy normally needs to be evaluated before being applied in the real design.In the last decade, prior researchers prefer to utilize the cycle-accurate simulations or trace-driven simulations to study the prefetching behaviors.However, as the increasing complexity of hardware components, the huge time-consuming simulationbased methods would never be appropriate for performance evaluations.This paper proposes a method of modeling prefetching influences on cache misses using artificial neural networks, which has an average error of 8% compared to gem5 cycle-accurate simulations, and the performance prediction process can be sped up by 30 times on average.