Bandit-based Run-time Adaptation of Cache Replacement Policies in Content Management Systems
| Autor: | Srivattsan Sridharan, Shravan Gaonkar, José A. B. Fortes |
|---|---|
| Rok vydání: | 2020 |
| Předmět: |
Hardware_MEMORYSTRUCTURES
Computer science Distributed computing Rank (computer programming) 020206 networking & telecommunications 02 engineering and technology Web traffic Metric (mathematics) 0202 electrical engineering electronic engineering information engineering Reinforcement learning Cache Latency (engineering) Baseline (configuration management) Adaptation (computer science) |
| Zdroj: | CANDAR |
| Popis: | Web traffic is increasingly complex, in terms of both content accessed and geographical distribution of requests. Content providers usually employ geographically distributed web-caches to reduce latency in delivering content. A web-cache configured with a single replacement policy cannot provide optimal performance due to the inherent variability associated with the request pattern. We describe an adaptation mechanism (DETACH) that utilizes a multi-armed bandit approach to adapt the replacement policy depending on a new metric, the rank of evicted objects, which indirectly characterizes run-time request patterns. The proposed mechanism requires very minimal modifications to existing cache systems and does not maintain any history (ghost entries). We also describe the implementation of a cache system that uses the proposed mechanism to adapt the replacement policy at runtime. We consider two replacement policies, LRU and LFU, but the system is expandable to other policies. Our experiments with various datasets show that the system adapts quickly to the best performing replacement policy. We compare the performance of the system to a recently proposed approach called LeCaR, and baseline LRU and LFU policies. Of the 7 datasets used on our experiments, DETACH outperformed LeCaR in 4 of them by up to a factor of two. In the remaining 3 datasets LeCaR outperformed DETACH by up to 5%. Our proposed system consistently outperforms the baseline LRU and LFU replacement policies for all the datasets. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|