Energy-oriented dynamic SPM allocation based on time-slotted Cache conflict graph

Abstract

Energy consumption has always been considered as the key issue of the state-of-the-art SoCs. Implementing an on-chip Cache is one of the most promising solutions. However, traditional Cache may suffer from performance and energy penalties due to the Cache conflict. In order to deal with this problem, this paper firstly introduces a Time-Slotted Cache Conflict Graph to model the behavior of Data Cache conflict. Then, we implement an Integer Nonlinear Programming to select the most profitable data pages and employ Virtual Memory System to remap those data pages, which can cause severe Cache conflict within a time slot, to the on-chip Scratchpad Memory (SPM). In order to minimize the swapping overhead of dynamic SPM allocation, we introduce a novel SPM controller with a tightly coupled DMA to issue the swapping operations without CPU's intervention. The proposed method can optimize all of the data segments, including global data, heap and stack data in general, and reduce 24.83% energy consumption on average without any performance degradation.