Improving the Performance of STT-MRAM LLC through Enhanced Cache Replacement Policy

Pierre-Yves Péneau 1 David Novo 1 Florent Bruguier 1 Lionel Torres 1 Gilles Sassatelli 1 Abdoulaye Gamatié 1
1 ADAC - ADAptive Computing
LIRMM - Laboratoire d'Informatique de Robotique et de Microélectronique de Montpellier
Abstract : Modern architectures adopt large on-chip cache memory hierarchies with more than two levels. While this improves performance, it has a certain cost in area and power consumption. In this paper, we consider an emerging non volatile memory technology, namely the Spin-Transfer Torque Magnetic RAM (STT-MRAM), with a powerful cache replacement policy in order to design an efficient STT-MRAM Last-Level Cache (LLC) in terms of performance. Well-known benefits of STT-MRAM are their near-zero static power and high density compared to volatile memories. Nonetheless, their high write latency may be detrimental to system performance. In order to mitigate this issue, we combine STT-MRAM with a recent cache replacement policy. The benefit of this combination is evaluated through experiments on SPEC CPU2006 benchmark suite, showing performance improvements of up to 10% compared to SRAM cache with LRU on a single core system.
Document type :
Conference papers
Complete list of metadatas
https://hal-lirmm.ccsd.cnrs.fr/lirmm-01669254
Contributor : Pierre-Yves Péneau <>
Submitted on : Tuesday, October 2, 2018 - 11:13:28 AM
Last modification on : Tuesday, December 18, 2018 - 3:32:13 PM
Long-term archiving on : Thursday, January 3, 2019 - 1:41:32 PM

File

peneau2018improving.pdf
Files produced by the author(s)

Identifiers

Collections

LIRMM | LIRMM_MIC | MIC | MIPS | UNIV-MONTPELLIER | ADAC

Citation

Pierre-Yves Péneau, David Novo, Florent Bruguier, Lionel Torres, Gilles Sassatelli, et al.. Improving the Performance of STT-MRAM LLC through Enhanced Cache Replacement Policy. ARCS: Architecture of Computing Systems, Apr 2018, Braunschweig, Germany. pp.168-180, ⟨10.1007/978-3-319-77610-1_13⟩. ⟨lirmm-01669254v2⟩

Export

BibTeX TEI DC DCterms EndNote

Share

Metrics

Record views

88

Files downloads

133

Contact

support.ccsd.cnrs.fr
hal.support@ccsd.cnrs.fr
Logo CCSD