Designing efficient memory schedulers for future systems

Update Item Information
Publication Type dissertation
School or College College of Engineering
Department Computing
Author Chatterjee, Niladrish
Title Designing efficient memory schedulers for future systems
Date 2013-12
Description The internet-based information infrastructure that has powered the growth of modern personal/mobile computing is composed of powerful, warehouse-scale computers or datacenters. These heavily subscribed datacenters perform data-processing jobs under intense quality of service guarantees. Further, high-performance compute platforms are being used to model and analyze increasingly complex scientific problems and natural phenomena. To ensure that the high-performance needs of these machines are met, it is necessary to increase the efficiency of the memory system that supplies data to the processing cores. Many of the microarchitectural innovations that were designed to scale the memory wall (e.g., out-of-order instruction execution, on-chip caches) are being rendered less effective due to several emerging trends (e.g., increased emphasis on energy consumption, limited access locality). This motivates the optimization of the main memory system itself. The key to an efficient main memory system is the memory controller. In particular, the scheduling algorithm in the memory controller greatly influences its performance. This dissertation explores this hypothesis in several contexts. It develops tools to better understand memory scheduling and develops scheduling innovations for CPUs and GPUs. We propose novel memory scheduling techniques that are strongly aware of the access patterns of the clients as well as the microarchitecture of the memory device. Based on these, we present (i) a Dynamic Random Access Memory (DRAM) chip microarchitecture optimized for reducing write-induced slowdown, (ii) a memory scheduling algorithm that exploits these features, (iii) several memory scheduling algorithms to reduce the memory-related stall experienced by irregular General Purpose Graphics Processing Unit (GPGPU) applications, and (iv) the Utah Simulated Memory Module (USIMM), a detailed, validated simulator for DRAM main memory that we use for analyzing and proposing scheduler algorithms.
Type Text
Publisher University of Utah
Subject DRAM; Memory system; Scheduling
Dissertation Institution University of Utah
Dissertation Name Doctor of Philosophy
Language eng
Rights Management Copyright © Niladrish Chatterjee 2013
Format Medium application/pdf
Format Extent 2,756,257 bytes
Identifier etd3/id/2678
ARK ark:/87278/s6rv3wwh
Setname ir_etd
ID 196253
Reference URL https://collections.lib.utah.edu/ark:/87278/s6rv3wwh