Back to search

Cornell University

Criticality-Aware Memory Systems

Abstract

dc:description.abstract

Research on computer memory systems has been of increasing importance over the last decade, as they have become a significant bottleneck for application performance. While newer memory systems offer increased memory level parallelism, they cannot be used blindly due to contention for shared resources, making a beneficial and valid sequencing of memory requests requisite in order to exploit these improvements. Traditional approaches to improving this sequencing rely on highly sophisticated memory systems, where significant amounts of inference are often required to make these sophisticated decisions. Unfortunately, this design philosophy may no longer be sustainable. For example, as memory clock frequencies continue to scale while processor frequencies remain stagnant, sophisticated memory controllers are already being squeezed out, forcing computer architects to revert to simpler designs. We use this as an opportunity to symbiotically involve the processor cores in the decision-making process, simultaneously offloading the complexity from these memory decision makers while extracting richer information on each memory operation. This work studies the concept of load criticality, where the processors themselves identify the loads which they believe to be most important. Using loads that block at the end of the processor pipeline as an indicator of criticality, we annotate these load block predictions onto memory requests, for use by various components in memory. Our research finds that even using small, sim- ple predictors for load criticality can offer comparable performance to complex state-of-the-art schedulers for both parallel applications and multiprogrammed workloads on a contemporary multicore system. This same predictor can be used to obtain significant performance improvements and energy savings when using hardware prefetchers. Ultimately, our criticality-aware design approach achieves the performance of traditionally-complex memory systems, and does so with trivial overheads that are attractive for future commercial adoption.

Degree

thesis:*
Name thesis:degree_name
Ph. D., Electrical Engineering
Level thesis:degree_level
Doctor of Philosophy
Discipline thesis:degree_discipline
Electrical Engineering
Grantor
Cornell University
Year dc:date.issued
2014

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Ghose, Saugata
Chair dc:contributor.chair
  • Martinez, Jose F.
Committee members dc:contributor.committeemember
  • Manohar, Rajit
  • Albonesi, David H.

Subjects

dc:subject × 3

Rights

Language dc:language.iso
en_US

Identifiers

dc:identifier.*
Dc Identifier Bibid
8793459

Chain of custody

source
Harvested from
Cornell University
Base URL
ecommons.cornell.edu/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Ghose, Saugata. Criticality-Aware Memory Systems. Doctor of Philosophy thesis, Cornell University, 2014. https://hdl.handle.net/1813/38922