Since the energy consumed by the computing system among the total produced electric power has been gradually increased, demands on low electric power computing dramatically grew. Data centers of IT enterprises such as cloud computing center possess the 1.5% of total energy consumption and it still doubles every five years. Nowadays, recent technology faces the challenges to reduce energy consumption because of structural limitation of DRAM which consumes about 30% of total energy consumption in data centers. Meanwhile, PRAM (Phase change memory), witch do not require charge current thanks to its structural property, appeared. However, it is hard to alter DRAM as a main memory because of its low performance, high write power, and write endurance limitation.
To overcome the weak point of PRAM, the research related to the hybrid model combining PRAM with DRAM has been performed. Previous works on hybrid memory mostly focused on changing the memory controller to distinguish the frequently used page and rarely used page and save this information through the page trace. Using these data, the pages allocation are decided whether DRAM or PRAM. However, this approach is hard to be applied on the marketplace because established memory controller architecture needs to be changed. Furthermore, additional space and time to save the information is needed. To solve this problem, some efforts were perfomed to decide allocation of pages by using the information gathered from OS. But still, it is hard to find a better solution because the data obtained by OS is limited.
In this thesis, we propose an Adaptive Page Grouping (APG) algorithm, which manages the hybrid PRAM-DRAM main memory to reduce energy consumption. We suggest the method of storing the pages access information without using additional space while making it reacheable by the operating system (OS). We allocate pages effectively and reduce the migration among them through the grouping of pages which has similar ...