Owicki & Agarwal's study

Goals, methodology, conditions, and metrics

• static SW schemes versus HW snoopy scheme

  • Base: inconsistent caches - upper limit for performance
  • No-cache: pages with shared data are not allowed to be cached (C.mmp, Wulf& Bell)
  • Software-flush: program analysis and insertion of instructions for cache maintenance (Cytron et al.)
  • Dragon: HW scheme (snooping and write-broadcasting)

• MVA modeling

  • system model - number of cycles required for each operation
  • workload model - frequencies of specific actions
  • contention model - waiting caused by the contention on ICN

• parameter values: low, typical, and high

• performance measure: processing power