Retrieval-Augmented Generation (RAG) systems enhance Large Language Models (LLMs) by integrating external knowledge retrieval, but this integration increases query latency and cloud costs. Effective caching is critical to mitigate these overheads, especially in large-scale multi-tenant deployments where balancing efficiency, fairness, and cost is challenging. This paper presents a comprehensive simulation study comparing five strategies: global Least Frequently Used (LFU), global Least Recently Used (LRU), Window-LFU (LFU with aging), LFU+Quota (a soft per-tenant quota), and tenant-partitioned caching under realistic skewed workloads (Zipf \( {\user2{\alpha = 1}}{\mathbf{.0}} \) ) across 10,000 tenants. Fairness is quantified using Jain’s index (J) and a max-min proxy (minimum per-tenant hit rate \( {\mathbf{H}}_{{\min }} \) ). Our results demonstrate that global LFU achieves the highest cache hit rate (74.26%) and near-perfect fairness (Jain’s \( {\mathbf{J = 0}}{\mathbf{.995}} \) ), outperforming LRU (72.78% hit rate, \( {\mathbf{J = 0}}{\mathbf{.984}} \) ) and vastly surpassing tenant-partitioned caching (3.7% hit rate, \( {\mathbf{J = 0}}{\mathbf{.899}} \) ). Window-LFU achieves comparable efficiency (73.90% hit rate) with sustained fairness ( \( {\mathbf{J = 0}}{\mathbf{.991}} \) ), while LFU+Quota provides near-global efficiency (73.80% hit rate) with stronger tenant protections (higher \( {\mathbf{H}}_{{\min }} \) , \( {\mathbf{J}} \approx {\mathbf{0}}{\mathbf{.998}} \) ). The inherent workload skew enables frequency-aware policies to dominate, while strict partitioning leads to severe fragmentation and poor reuse. These findings highlight that frequency-based caches (LFU and Window-LFU) remain most efficient under skewed workloads, and LFU+Quota offers a practical efficiency–fairness trade-off for multi-tenant GenAI services.
