Power reduction controll for multicores in OSCAR multigrain parallelizing compiler
Jun Shirako, Keiji Richard Kimura, Hironori Kasahara · 2008
Multicore processors have become mainstream computer architecture to go beyond the performance and power efficiency limits of single-core processors. To achieve low power consumption and high performance on multicores, parallelizing compilers take on an important role. This paper describes the performance of a compiler-based power reduction scheme cooperating with OSCAR multigrain parallelizing compiler on a newly developed 8-way SH4A low power multicore chip for consumer electronics, which supports DVFS (dynamic voltage and frequency scaling) and clock/power gating. Using hardware parameters and parallelized program information, OSCAR compiler determines suitable voltage and frequency of each active processor core and appropriate schedule of clock gating and power gating. Performance experiments shows the compiler reduces consumed power by 88.3%, namely from 5.68 W to 0.67 W, for real-time secure AAC Encoding and 73.5%, namely from 5.73 W to 1.52 W, for real-time MPEG2 Decoding on 8 core execution.