A New Case for Skewed-Associativity

Skewed-associative caches have been shown to statisticaly exhibit lower miss ratios than set-associative caches at equal associativity degrees. But for L1 caches, the performance of a four-way set-associative cache is known to be quite close to that of a 16-way set-associative cache. In this paper, we rst show that BTBs and L2 caches featuring long line size may beneet a lot from a high associativity degree on set-associative caches and a LRU replacement policy. Skewed-associative caches with an hypothetic LRU replacement policy \provide" this high associativity but with a low number of cache banks. Unfortunately, LRU replacement policies can not be easily implemented in hardware on skewed-associative caches. We propose some new replacement heuristics for skewed-associative caches. Those heuristics exhibit slightly worse behavior than a full LRU replacement policy, but still allow a 2-way skewed-associative cache to outperform a 4-way LRU set-associative cache and a 4-way skewed-associative cache to outperform a 16-way LRU set-associative cache. Therefore skewed-associativity allows signiicant miss reduction on BTBs and L2 caches. Une nouvelle argumentation pour les structures associatives brouill ees R esum e : Il a et e montr e que les caches associatifs brouill es exhibent de meilleurs taux de succ es que les caches associatifs par ensemble. Dans ce rapport, nous montrons que les structures associatives brouill ees sont particuli e-rement int eressantes pour les caches de pr edictions de branchement et les caches secondaires. De plus nous proposons de nouvelles politiques de remplacement pour les structures asso-ciatives brouill ees Mots-cl e : caches associatifs brouill es 3 Abstract Skewed-associative caches have been shown to statisticaly exhibit lower miss ratios than set-associative caches at equal associativity degrees. But for L1 caches, the performance of a four-way set-associative cache is known to be quite close to that of a 16-way set-associative cache. In this paper, we rst show that BTBs and L2 caches featuring long line size may beneet a lot from a high associativity degree on set-associative caches and a LRU replacement policy. Skewed-associative caches with an hypothetic LRU replacement policy \provide" this high associativity but with a low number of cache banks. Unfortunately, LRU replacement policies can not be easily implemented in hardware on skewed-associative caches. We propose some new replacement heuristics for skewed-associative caches. Those heuristics exhibit slightly worse behavior than a full LRU replacement policy, but still allow a 2-way skewed-associative cache to outperform a 4-way LRU set-associative cache and …