Selectively Breaking Data Dependences to Improve the Utilization of Idle Cycles in Algorithm Level Re-Computing Data Paths

Kaijie Wu, Ramesh Karri

Research output: Contribution to journalArticle

Abstract

Although algorithm level re-computing techniques can trade-off the fault detection capability vs. time overhead of a Concurrent Error Detection (CED) scheme, they result in 100% time overhead when the strongest CED capability is achieved. Using the idle cycles in the data path to do the re-computation can reduce this time overhead. However, dependences between operations prevent the re-computation from fully utilizing the idle cycles. Deliberately breaking some of these data dependences can further reduce the time overhead associated with algorithm level re-computing. According to the experimental results the proposed technique, it brings time overhead down to 0-60% while the associated hardware overhead is from 12% to 50% depending on the design size.

Original languageEnglish (US)
Pages (from-to)501-511
Number of pages11
JournalIEEE Transactions on Reliability
Volume52
Issue number4
DOIs
StatePublished - Dec 2003

Fingerprint

Error detection
Fault detection
Hardware

Keywords

  • Concurrent error detection
  • Radiation hardening
  • Very deep sub-micron device

ASJC Scopus subject areas

  • Electrical and Electronic Engineering
  • Hardware and Architecture
  • Computer Graphics and Computer-Aided Design
  • Software
  • Safety, Risk, Reliability and Quality

Cite this

@article{2652d81b38a8404cb5e7b2b2145bfa56,
title = "Selectively Breaking Data Dependences to Improve the Utilization of Idle Cycles in Algorithm Level Re-Computing Data Paths",
abstract = "Although algorithm level re-computing techniques can trade-off the fault detection capability vs. time overhead of a Concurrent Error Detection (CED) scheme, they result in 100{\%} time overhead when the strongest CED capability is achieved. Using the idle cycles in the data path to do the re-computation can reduce this time overhead. However, dependences between operations prevent the re-computation from fully utilizing the idle cycles. Deliberately breaking some of these data dependences can further reduce the time overhead associated with algorithm level re-computing. According to the experimental results the proposed technique, it brings time overhead down to 0-60{\%} while the associated hardware overhead is from 12{\%} to 50{\%} depending on the design size.",
keywords = "Concurrent error detection, Radiation hardening, Very deep sub-micron device",
author = "Kaijie Wu and Ramesh Karri",
year = "2003",
month = "12",
doi = "10.1109/TR.2003.821942",
language = "English (US)",
volume = "52",
pages = "501--511",
journal = "IEEE Transactions on Reliability",
issn = "0018-9529",
publisher = "Institute of Electrical and Electronics Engineers Inc.",
number = "4",

}

TY - JOUR

T1 - Selectively Breaking Data Dependences to Improve the Utilization of Idle Cycles in Algorithm Level Re-Computing Data Paths

AU - Wu, Kaijie

AU - Karri, Ramesh

PY - 2003/12

Y1 - 2003/12

N2 - Although algorithm level re-computing techniques can trade-off the fault detection capability vs. time overhead of a Concurrent Error Detection (CED) scheme, they result in 100% time overhead when the strongest CED capability is achieved. Using the idle cycles in the data path to do the re-computation can reduce this time overhead. However, dependences between operations prevent the re-computation from fully utilizing the idle cycles. Deliberately breaking some of these data dependences can further reduce the time overhead associated with algorithm level re-computing. According to the experimental results the proposed technique, it brings time overhead down to 0-60% while the associated hardware overhead is from 12% to 50% depending on the design size.

AB - Although algorithm level re-computing techniques can trade-off the fault detection capability vs. time overhead of a Concurrent Error Detection (CED) scheme, they result in 100% time overhead when the strongest CED capability is achieved. Using the idle cycles in the data path to do the re-computation can reduce this time overhead. However, dependences between operations prevent the re-computation from fully utilizing the idle cycles. Deliberately breaking some of these data dependences can further reduce the time overhead associated with algorithm level re-computing. According to the experimental results the proposed technique, it brings time overhead down to 0-60% while the associated hardware overhead is from 12% to 50% depending on the design size.

KW - Concurrent error detection

KW - Radiation hardening

KW - Very deep sub-micron device

UR - http://www.scopus.com/inward/record.url?scp=0842333932&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=0842333932&partnerID=8YFLogxK

U2 - 10.1109/TR.2003.821942

DO - 10.1109/TR.2003.821942

M3 - Article

VL - 52

SP - 501

EP - 511

JO - IEEE Transactions on Reliability

JF - IEEE Transactions on Reliability

SN - 0018-9529

IS - 4

ER -