Cache Design Space Exploration For Android Applications

With the rapid improvments of CPUs, Caches have been widely used to solve the problem of the performance gap between processors and memory. Due to the ever enlarging market share of Android, and the difference of memory accessing paterns between Android appliacations and traditional desktop applications shown by some previous studies, it is very meaningful to reaserch cache design space exploration in the Android scenarios.The mian work of this thesis has two parts, the first is establishing a cache model for out-of-order processors, the second is exploring the design space of Cache in Android scenarios based on the model we proposed. The stack distance theory is an important way to reasearch Caches. However, out-of-order processors have an influence on the stack distance distribution, thus the cache model based on stack distance theroy can not be applied directly on out-of-order processors. Neural networks is applied in this thesis to improve the cache model based on stack distance, its basic thought is using artificial neural network to fit the changes of the stack distance distribution. The model uses memory trace to predict Cache miss ratio. Compared to Gem5 full simulations, the model has about 5% error rate, while reducing the simulation time by about 80%.Using our cache model and Cacti, this thesis analyzes the influence on cache performance, area and power caused by Cache capacity, associativity and line size, and finds an optimal configuration of Cache for Android applications.