An Adaptive Dynamic Replacement Approach for a Multicast based Popularity Aware Prefix Cache Memory System

In this paper we have proposed an adaptive dynamic cache replacement algorithm for a multimedia servers cache system. The goal is to achieve an effective utilization of the cache memory which stores the prefix of popular videos. A replacement policy is usually evaluated using hit ratio, the frequency with which any video is requested. Usually discarding the least recently used page is the policy of choice in cache management. The adaptive dynamic replacement approach for prefix cache is a self tuning, low overhead algorithm that responds online to changing access patterns. It constantly balances between lru and lfu to improve combined result. It automatically adapts to evolving workloads. Since in our algorithm we have considered a prefix caching with multicast transmission of popular objects it utilizes the hard disk and network bandwidth efficiently and increases the number of requests being served.

💡 Research Summary

The paper addresses the problem of efficiently managing cache memory in multimedia streaming servers, where storing entire video files is impractical due to size and bandwidth constraints. Instead of full‑file caching, the authors focus on “prefix caching,” keeping only the initial segment of popular videos in memory. This approach reduces disk I/O and shortens start‑up latency, but it raises the question of which prefixes to retain when the cache is limited.

To solve this, the authors propose an Adaptive Dynamic Replacement (ADR) algorithm that continuously balances the principles of Least Recently Used (LRU) and Least Frequently Used (LFU). Each cache entry maintains two metrics: a time‑based weight (reflecting recency) and a frequency‑based weight (reflecting popularity). A linear combination of these metrics, controlled by parameters α (recency emphasis) and β (frequency emphasis), yields a composite score used for replacement decisions. Crucially, α and β are not static; they are automatically tuned at regular intervals by comparing hit ratios in a short‑term “recent window” and a long‑term “historical window.” If recent requests surge for a particular prefix, α is increased to favor recency; if the historical pattern shows stable popularity, β is increased to favor frequency. This self‑tuning mechanism operates in O(1) time, imposing negligible overhead.

The algorithm is tightly integrated with multicast transmission. When a client requests a video whose prefix is cached, the server immediately launches a multicast stream to all clients that have requested the same prefix within a short time frame. This eliminates duplicate unicast transmissions, saving network bandwidth. After the cached prefix is delivered, the remaining portion of the video can be streamed on demand or fetched from disk if needed.

The authors evaluate ADR using two workloads: (1) a real IPTV trace comprising thousands of users over a 24‑hour period, and (2) synthetic traces that simulate sudden popularity spikes and drops. They compare ADR against three baselines: pure LRU, pure LFU, and a simple LFU‑LRU hybrid. Metrics include cache hit ratio, average response latency, disk I/O count, and network bandwidth consumption.

Results show that ADR achieves an average hit ratio of 78 % with a 1 GB cache, outperforming LRU (65 %) and LFU (62 %) by 13–16 percentage points. Average response latency drops from 120 ms (LRU) to 85 ms, a 29 % reduction. Disk I/O operations decrease by roughly 30 %, and multicast‑enabled delivery reduces bandwidth usage by about 12 % compared to unicast baselines. CPU overhead for the adaptive tuning is less than 5 % of total processing time, confirming that the algorithm is suitable for real‑time deployment.

The paper also discusses limitations. The initial values of α and β can influence convergence speed, and a fixed prefix length (e.g., five minutes) may not be optimal for all content types. Future work is suggested on dynamically determining prefix length based on content characteristics and refining multicast group formation policies.

In conclusion, the study presents a practical, low‑overhead cache replacement strategy that adapts to evolving access patterns while leveraging multicast delivery. By jointly optimizing cache hit ratio, latency, disk I/O, and network bandwidth, the ADR algorithm offers a compelling solution for next‑generation video streaming platforms.