Also as “ARC: A Self-Tuning, Low Overhead Replacement Cache” by the same authors. In Proceedings of 2nd USENIX Conference on File and Storage Technologies (FAST), pp.115–130, San Francisco CA, Mar 31–Apr 2 2003.

This is the paper that proposed Adaptive Replacement Cache (ARC) as a caching algorithm.

ARC implements with two LRU (lease recently used) queues $L_1$ and $L_2$. Where $L_1$ is storing objects that seen only once and $L_2$ is storing objects that referenced at least twice. Supposedly, the union of these two LRU queues, which each holds at most $c$ items, can provide a better cache than a single LRU queue because it solves the temporal locality problem. But then the total amount of memory is $2c$.

ARC therefore assumes $L_1=T_1+B_1$ and $L_2=T_2+B_2$ which the union of $L_1$ and $L_2$ is called the directory (with size limit of $2c$ and $L_1$ is no bigger than $c$) and the union of $T_1$ and $T_2$ is the cache (with size limit of $c$). Only the items in cache is stored in memory. Directory indexes the ID of the objects only. The maximum size of the caches $T_1$ and $T_2$ are $p$ and $c-p$ respectively, where $p$ is adaptive.

When an object is referenced, it is put into the cache, using the following algorithm:

  • If cache hit in $T_1+T_2$: Move the item to MRU position of $T_2$
  • If directory hit in $B_1$, then p is too small. So load the object into MRU position of $T_2$ and set $p=\min(c, p+\max(\lvert B_2\rvert/\lvert B_1\rvert),1))$
  • If directory hit in $B_2$, then p is too big. So load the object into MRU position of $T_2$ and set $p=\min(c, p-\max(\lvert B_1\rvert/\lvert B_2\rvert),1))$
  • If directory miss, do not adjust p but replace items in the directory/cache as follows
    • If $\lvert L_1\rvert=c$, delete the LRU of $L_1$ (which is also the LRU of $B_1$)
    • If $\lvert L_1\rvert<c$, delete the LRU of $L_2$ (which is also the LRU of $B_2$)
    • Then load the item into the MRU position of $T_1$

In the above algorithm, we will move an item from $T$ to $B$ if necessary so that the size constraint imposed by the parameter $p$ is honoured.

Bibliographic data

   title = "Outperforming LRU with an Adaptive Replacement Cache Algorithm",
   author = "Nimrod Megiddo and Dharmendra S. Modha",
   affiliation = "IBM Almaden Research Center",
   journal = "IEEE Computer Magazine",
   pages = "58--65",
   month = "April",
   year = "2004",