一种基于访问计数的SPM管理策略

摘要

Traditional Scratch Pad Memory(SPM) management depends on compiler, thus reducing the portability and ability of cross-platform deployment. A SPM management strategy is proposed based on access counting to guide the SPM allocation without compiler support. The approach predicts the application’s core work sets at runtime by counting the memory references. Experimental results show that the new approach reduces the energy 32.5%while keeping program’s stable running compared with Cache systems.%传统便签存储器(SPM)管理方法依赖于编译器提供支持，不利于应用程序的移植和跨平台部署，且难以反映程序动态运行特征。为此，提出一种基于访问计数的SPM动态管理策略方法，通过在程序运行时对访存地址进行访问计数，预测程序核心工作集并指导SPM动态分配。实验结果表明，与单纯Cache系统相比，该方法在保持程序执行性能稳定的基础上，能耗降低了约32.5%。