Daniel Berend

We introduce the AdaptiveClimb cache policy, a new policy for cache management. This policy improves LRU to gain higher performance in a fixed probability scenario, without maintaining statistics for each item. Rather, it stores a single value (the current jump), while preserving the fast adaptation of probability changes of LRU. AdaptiveClimb is a modification of CLIMB cache policy, that unlike CLIMB changes the number of positions shifts according to whether the last request has been a hit or a miss. Performance of CLIMB is close to that of the optimal off-line algorithm, but its stabilization time is long. LRU is much more sensitive to changes, but it is sensitive to noise. Thus, AdaptiveClimb combines these two advantages in a single algorithm with both good performance and short stabilization time.