Finding Energy Sustainable Techniques for Computer Memory

Abstract

The increasing growth of internet based information infrastructure consists of powerful computers and data centers has led to the development of personal/mobile computing devices. These datacenters are required to perform heavy data processing while ensuring quality of service. These highly powerful performance oriented computing platforms are used to model and analyze various complicated scientific problems and natural phenomenon. These computing devices are capable of providing high performance requirements of various complex applications only if they are facilitated with efficient memory system. The memory system stores data during and after execution and provides data to processing cores necessary to complete execution of applications. It is main memory system that stores data structure necessary for completion of a program. Main memory constitutes a major part in overall system s power consumption. Researchers have reported that in mid-range IBM eServer machine, main memory contributed 40% of the total system s power consumption. newlineMany researchers have worked to make DRAM based main memory system efficient in terms of performance, energy and access locality. The prime factor that affects the efficiency of main memory system is its memory controller. Memory accesses constitute an important part of total applications energy consumption. It is memory controller of main memory system that is responsible for its efficient working. The decision regarding what command to issue and when, is newlinedependent on main memory controller. Main memory controller makes this decision on the basis of memory access scheduler used. Based on utilized memory scheduling policy DRAM requests are serviced, the sequence of servicing requests targeted to Dynamic-RAM largely affects the performance as well as energy consumed by main memory sub-system. This thesis investigates the impact of memory scheduling policy on energy consumption and performance of main memory system in several situations. We propose four scheduling policies that try to