Energy oriented optimization of dynamic management strategy for scratch-pad memory

Effective utilization of on-chip storage space is important to reduction of memory subsystem energy consumptions of embedded systems which are powered by batteries. On-chip cache memory helps to alleviate this bottleneck to some extent and has been widely used in the past. However, two factors, invalid transmission between on-chip memory and off-chip memory, limited effectiveness of compiler, both indicate that it may not be the best candidate for portable/embedded devices. Software-managed on-chip memory, which generally is called as scratch-pad memory (SPM) may be a better alternative for exploiting locality, meanwhile, how to efficiently map code & date into SPM is still a problem. In this paper, we propose and evaluate a compiler-controlled dynamic SPM management strategy which analyzes the application and inserts instructions to dynamically copy both code segments and variables onto SPM at runtime. Compared with previous work, this strategy is dynamic and it allows the change of the contents of the SPM during execution