• DocumentCode
    2136387
  • Title

    Variable-based multi-module data caches for clustered VLIW processors

  • Author

    Gibert, Enric ; Abella, Jaume ; Sánchez, Jesús ; Vera, Xavier ; González, Antonio

  • Author_Institution
    Departament d´´Arquitectura de Computadors, Univ. Politecnica de Catalunya, Barcelona, Spain
  • fYear
    2005
  • fDate
    17-21 Sept. 2005
  • Firstpage
    207
  • Lastpage
    217
  • Abstract
    Memory structures consume an important fraction of the total processor energy. One solution to reduce the energy consumed by cache memories consists of reducing their supply voltage and/or increase their threshold voltage at an expense in access time. We propose to divide the L1 data cache into two cache modules for a clustered VLIW processor consisting of two clusters. Such division is done on a variable basis so that the address of a datum determines its location. Each cache module is assigned to a cluster and can be set up as a fast power-hungry module or as a slow power-aware module. We also present compiler techniques in order to distribute variables between the two cache modules and generate code accordingly. We have explored several cache configurations using the Mediabench suite and we have observed that the best distributed cache organization outperforms traditional cache organizations by 19%-31% in energy-delay and by 11%-29% in energy-delay. In addition, we also explore a reconfigurable distributed cache, where the cache can be reconfigured on a context switch. This reconfigurable scheme further outperforms the best previous distributed organization by 3%-4%.
  • Keywords
    cache storage; memory architecture; multiprocessing systems; parallel architectures; program compilers; L1 data cache; Mediabench suite; cache memories; clustered VLIW processors; compiler techniques; distributed cache organization; memory structures; reconfigurable distributed cache; supply voltage; threshold voltage; variable-based multi-module data caches; Computer architecture; Concurrent computing; Delay; Global communication; Parallel architectures; Processor scheduling; Registers; Runtime; VLIW; Wire;
  • fLanguage
    English
  • Publisher
    ieee
  • Conference_Titel
    Parallel Architectures and Compilation Techniques, 2005. PACT 2005. 14th International Conference on
  • ISSN
    1089-795X
  • Print_ISBN
    0-7695-2429-X
  • Type

    conf

  • DOI
    10.1109/PACT.2005.40
  • Filename
    1515594