Hierarchical Power Management Framework on Manycore Systems Using OS Migration Techniques

Abstract

Recently, power management schemes are required in manycore systems for gratifying the chip-level power and thermal constraints. Most research focuses on optimizing power consumption for parallel applications running on all cores. In this thesis, we tried to investigate the possibilities of feasible power man- agement of a manycore systems running several independent operating systems (OS). Exploiting hardware support such as memory address indirection and global shared memory, a zero-copy OS migration technique is implemented on the Linux platform with minimal overhead. In the context of dynamic voltage and frequency scaling for many-core chips where due to the cost and few practi- cal reasons, the voltage and frequency can only be controlled for a group of cores physically grouped into a voltage and/or frequency domain. In this case, Zero- copy OS migration can be employed to group OSes with similar performance characteristics into one voltage domain which then allows DVFS algorithm to better match the voltage and frequency settings to the characteristics of that domain. From hierarchical power management framework implemented on the Intel Single-chip Cloud Computer (SCC) running up to 40 independent Linux instances, We show that our approach, on average, saves the energy consump- tion by 30% over existing DVFS policies for a wide range of load patterns with minimal performance degradation. In the conclusion, it can improve the performance per watt ratio by 27% in most of benchmark situations.