Popularity Driven Caching for Static Caches Based on Partial Replicas

Abstract

Content-based Information retrieval system are gaining importance with the irruption of smartphones and online services. Fast response time and high traffic is crucial to provide a good service to users worldwide. In contrast with metadata-based systems, content-based systems deal with a larger amount of data. Therefore, the lookup approach and optimization techniques cannot be the same. We faced those difficulties to improve performance on content-based systems. This paper present a static cache based on partial replication for IR system in which popular items are requested often through queries with different con- tent. The static cache stores a subset of the data in order to reduce the search time. In addition, we present two popularity driven cache replacement algo- rithms that are able to predict which items will be probably popular in the near future to include them in cache. The first algorithm is DerPop that predicts popular items using derivatives to estimate how much will increase or decrease the number of request for each song in the future. The other algorithm is MixPop that combines LRU and DerPop. These two algorithms outperformed LRU by 132% and 145% in terms of cache hits. Experiments were carried out iin Echoprint utilizing a new benchmark for audio identification that models userss listing patterns to provide a useful test workload.