Towards a predictive cache replacement strategy for multimedia content

Abstract

In recent years, telecom operators have been moving away from traditional broadcast-driven television, towards IP-based interactive and on-demand multimedia services. Consequently, multicast is no longer sufficient to limit the amount of generated traffic in the network. In order to prevent an explosive growth in traffic, caches can be strategically placed throughout the content delivery infrastructure. As the size of caches is usually limited to only a small fraction of the total size of all content items, it is important to accurately predict future content popularity. Traditional caching strategies only take into account the past when deciding what content to cache. Recently, a trend towards novel strategies that actually try to predict future content popularity has arisen. In this article, we ascertain the viability of using popularity prediction in realistic multimedia content caching scenarios. The proposed generic popularity prediction algorithm is capable of predicting future content popularity, independent of specific content and service characteristics. Additionally, a novel cache replacement strategy, which employs the popularity prediction algorithm when making its decisions, is introduced. A detailed evaluation, based on simulation results using trace files from an actual deployed Video on Demand service, was performed. The evaluation results are used to determine the merits of popularity-based caching compared to traditional strategies. Additionally, the synergy between several parameters, such as cache size and prediction window, is investigated. Results show that the proposed prediction-based caching strategy has the potential to significantly outperform state-of-the-art traditional strategies. Specifically, the evaluated Video on Demand scenario showed a performance increase of up to 20% in terms of cache hit rate.

A survey on predicting the popularity of web content

“…One way to improve the efficiency of cache replacement algorithms is to actually integrate popularity prediction methods in the cache replacement decision. Famaey et al propose P-LFU, an adaptation of LFU that determines which content to evict from the cache based on the predicted future demand [102]. Four generic functions (linear, power-law, exponential, and Gaussian) have been used to predict future content demand with the exponential distribution showing the most accurate results.…”

Challenges to support edge-as-a-service

“…These works typically identify correlations between specific video characteristics and future popularity. Famaey et al [12] found that analytical models can be used to predict the distribution associated with specific popularity of video contents. These problems are highly relevant to the EaaS problem space, as mobile users are access higher volumes of mobile data and video content.…”

A novel caching algorithm for VoD proxy implementation and its evaluation including a new set of metrics for efficiency analysis

“…Many caching algorithms use the history of the access pattern of the clients as a reference to try to predict which portions of the collection will be needed in the future [6][7][8][9][10][11]. Thus, the most accessed video objects should be stored in the cache in order to easily serve future requests.…”