On Selective Placement for Uniform Cache Objects

Abstract

For several decades, numerous studies for the cache replacement algorithms have been performed in order to enhance cache performance. But, as the performance of a new algorithm has improved, the complexity of the algorithm has gradually increased. In the end, in most of practices, the LRU algorithm has been still successfully used, because of its good hit ratio, low complexity of O(1), and so on.

For most of cache replacement algorithms, all references are always cached without any limitation. But, in the reference stream which follows the well-known Zipf-like nature of popularity, such as Web server/proxy and P2P file sharing environments, a large fraction of the total reference stream are arbitrarily singly or rarely accessed objects. These unpopular objects tend to cause unnecessary replacement of popular ones. Thus, to account for this, a mechanism such as the cache placement policy needs to be considered.

In this paper, we proposed the recency-based cache placement strategy to effectively capture the popularity of a newly referenced object, and showed the superiority of the cache placement policy by performance comparison with simulation experiments. Finally we concluded that simply combining with the cache placement policy enables to raise the hit ratio without need of complex replacement algorithms.