Adverse Filtering Effects and the Resilience of Aggregating Caches

Abstract

Cache hit rates can be dramatically affected by the pre-filtering effects of successive cache stages. Where the intervening cache is of comparable size to the target cache, this effect can render traditional replacement policies totally ineffective. We propose that such scenarios are increasingly common in distributed and mobile environments. We present a study of the effects of intervening cache filtering on hit rates, demonstrating the rapid degradation of performance for traditional cache replacement schemes. We also compare the performance of a mechanism we have recently developed, the aggregating cache. The aggregating cache exhibits tremendously improved performance degradation under even the most problematic scenarios.