ARC: A SELF-TUNING, LOW OVERHEAD REPLACEMENT CACHE Nimrod Megiddo Dharmendra S. Modha IBM Almaden Research Center

Introduction (I) Caching is widely used in storage systems databases web servers processors file systems disk drives RAID controllers operating systems

Introduction (II) ARC is a new cache replacement policy: Scan-resistant: Better than LRU Self-tuning: Avoids problem of many recent cache replacement policies Tested on numerous workloads

Our Model (I) Cache/ Main Memory (pages) Secondary Storage (pages) on demand replacement policy Secondary Storage (pages)

Our Model (II) Caches stores uniformly sized items (pages) On demand fetches into cache Cache expulsions decided by cache replacement policy Performance metrics include Hit rate (= 1 - miss rate) Overhead of policy

Previous Work (I) Offline Optimal (MIN): replaces the page that has the greatest forward distance Requires knowledge of future Provides an upper-bound Recency (LRU): Most widely used policy Frequency (LFU): Optimal under independent reference model

Previous Work (II) LRU-2: replaces page with the least recent penultimate reference Better hit ratio Needs to maintain a priority queue Corrected in 2Q policy Must still decide how long a page that has only been accessed once should be kept in the cache 2Q policy has same problem

LRU expels B because A was accessed after this last reference to B Example Last two references to pages A and B X A X B X B X A Time LRU -2 expels A because B was accessed twice after this next to last reference to A LRU expels B because A was accessed after this last reference to B

Previous Work (III) Low Inter-Reference Recency Set (LIRS) Frequency-Based Replacement (FBR) Least Recently/Frequently Used(LRFU): subsumes LRU and LFU All require a tuning parameter Automatic LRFU (ALRFU) Adaptive version of LRFU Still requires a tuning parameter

ARC (I) Maintains two LRU lists Pages that have been referenced only once (L1) Pages that have been referenced at least twice (L2) Each list has same length c as cache Cache contains tops of both lists: T1 and T2 Bottoms B1 and B2 are not in cache

ARC (II) L-1 L-2 T1 T2 “Ghost caches” (not in memory) B1 B2 |T1| + |T2| = c “Ghost caches” (not in memory) B2

ARC (III) ARC attempts to maintain a target size target_T1 for list T1 When cache is full, ARC expels LRU page from T1 if |T1|  target_T1 LRU page from T2 otherwise

ARC (IV) If missing page was in bottom B1 of L-1, ARC increases target_T1 target_T1= min(target_T1+max(|B2|/|B1|,1),c) If missing page was in bottom B2 of L-2, ARC decreases target_T1 target_T1= max(target_T1-max(|B1|/|B2|,1),0)

ARC (V) Overall result is Two heuristics compete with each other Each heuristic gets rewarded any time it can show that adding more pages to its top list would have avoided a cache miss Note that ARC has no tunable parameter Cannot get it wrong!

Experimental Results Tested over 23 traces: Always outperforms LRU Performs as well as more sophisticated policies even when they are specifically tuned for the workload Sole exception is 2Q Still outperforms 2Q when 2Q has no advance knowledge of the workload characteristics.