Optimization of Data Allocation on CMP Embedded System with Data Migration

Abstract

Chip multi-processors are applied in embedded system. An embedded system with multi-cores is considered large and consumes substantial power. Scathed-pad memory (SPM) and non-volatile memory (NVM) are new memory technologies, and an embedded system that uses SPM and NVM can reduce its size and power consumption.This study proposes an optimization of data allocation with data migration algorithm on task-level (TODMA). Data migration and dynamic programming are co-dependent and are combined to allocate task data in TODMA. In the experiments, we evaluated the performance of TODMA algorithm based on DSPstone benchmark and random benchmark. Results of DSPstone show that TODMA reduces the time cost, the number of write activities on NVM, and system energy consumption by 36.25, 24.58, and 34.41 %, respectively, compared with the greedy algorithm. The corresponding reductions are 33.82, 10.00, and 24.27 %, respectively, compared with the iterational optimal data placement algorithm (IODA). For the random benchmark, TODMA can reduce the time cost, the number of write activities on NVM, and system energy consumption by 26.79, 33.32, and 26.88 %, respectively, compared with the greedy algorithm. The advanced percentages are 25.17, 9.87, and 19.54 %, respectively, which are similar to IODA algorithm. Results show that the proposed TODMA algorithm effectively optimizes data allocation problems, improves system performance, reduces the number of write activities on NVM main memory, and lessens system energy consumption.