Regresar

An evaluation of cache management policies under workloads with malicious requests

Abstract:

We study the performance of cache admission and eviction policies under workloads with malicious requests (i.e., with requests that try to poison or pollute the cache with useless information). The eviction policies we have chosen for this study are: LRU, SLRU, and SLRU2. We also evaluated W-TinyLFU, a frequency-based cache admission policy that seeks to increase the effectiveness of the cache, specially under skewed access distributions. We implemented the policies in a caching simulator and evaluated them using Zipfian and real workloads from Yahoo, YouTube, and a feature animation company. In addition, we created modified versions of these workloads with added malicious requests, and use these modified workloads to observe how the malicious requests affect the performance of the policies under study. To the best of our knowledge, this is the first systematic study of the performance of caching policies under workloads with malicious requests. Our results show that W-TinyLFU improves the performance of eviction policies such as LRU and SLRU2 for both real and synthetic workloads, including those that contain malicious requests.