On the finding proper cache prediction model using neural network

Cache is an important component in computer architecture. It is a kind of memories that is located very close to CPU and can be accessed fast. With the current technology, the cache price is still expensive and thus the size cannot be very large. With the proper selection of the cache design, one can save the total hardware cost while certainly gaining the fast execution of programs. In this paper, we apply the data mining technique to find the proper cache model. We particularly consider to predict the cache block size. First, how to collect the address reference traces is described. Several tools are used to collect the traces. We are interested in the trace of the data mining benchmark, NUMineBench[4]. From the traces, the reference patterns are analyzed. Patterns related to block size are extracted. These features together with the trace behaviors from the simulation are used to build the prediction model, i.e, neural network. The methodology is found to be effective and can be expanded to consider other parameters such as cache capacity, associativity etc.