Sobre a confiabilidade da amostragem de dados para apoio à tomada de decisões de gerenciamento de sistemas de cache.

Abstract

Workload analysis is essential for the design and optimization of caching systems, guiding decisions such as capacity planning, the selection of replacement policies, and admission control policies, among others. This analysis is frequently based on traces from production systems. While it is ideal to analyze the entirety of requests received by a system, the computational overhead imposed by data collection often makes sampling a practical necessity. Despite its widespread use, the impact of sampling on workload representativeness has been largely overlooked. This work highlights these representativeness issues by showing that systematic sampling, a simple technique, can significantly distort workload characteristics, leading to misguided decisions. Specifically, our results reveal that sampled traces can cause substantial errors in estimating key metrics, such as item access frequency, producing results that diverge from those based on full traces. On the other hand, we found that for some workloads, another technique, block sampling, while imperfect, can improve sample representativeness by preserving important workload characteristics. These findings serve as a warning to researchers and practitioners who rely on sampled traces, highlighting the risks of applying sampling uncritically in the analysis of caching systems.