Deriving and visualizing the lower bounds of information gain for prefetch systems

摘要

While prefetching scheme has been used in different levels of computing, research works have not gone far beyond assuming a Markovian model and exploring localities in various applications. In this paper, we derive two lower bounds of information gain for prefetch systems and approximately visualize them in terms of decision tree learning concept. With the lower bounds of information gain, we can outline the minimum capacity required for a prefetch system to improve performance in respond to the probability model of a data set. By visualizing the analysis of information gain, We also conclude that performing entropy coding on the attibutes of a data set and making prefetching decisions based on the encoded attributes can help lowering the requirement of information tracking capacity.