Hybrid multi-tenant cache management for virtualized ISP networks

Abstract

In recent years, Internet Service Providers (ISPs) have started to deploy Telco Content Delivery Networks (Telco CDNs) to reduce the pressure on their network resources. The deployment of Telco CDNs results in reduced ISP bandwidth utilization and improved service quality by bringing the content closer to the end-users. Furthermore, virtualization of storage and networking resources can open up new business models by enabling the ISP to simultaneously lease its Telco CDN infrastructure to multiple third parties. Previous work has shown that multi-tenant proactive resource allocation and content placement can significantly reduce the load on the ISP network. However, the performance of this approach strongly depends on the prediction accuracy for future content requests. In this paper, a hybrid cache management approach is proposed where proactive content placement and traditional reactive caching strategies are combined. In this way, content placement and server selection can be optimized across tenants and users, based on predicted content popularity and the geographical distribution of requests, while simultaneously providing reactivity to unexpected changes in the request pattern. Based on a Video-on-Demand (VoD) production request trace, it is shown that the total hit ratio can be increased by 43% while using 5% less bandwidth compared to the traditional Least Recently Used (LRU) caching strategy. Furthermore, the proposed approach requires 39% less migration overhead compared to the proactive placement approach we previously proposed in Claeys et al. (2014b) and achieves a hit ratio increase of 19% and bandwidth usage reduction of 7% in the evaluated VoD scenarios and topology.

Self-Adaptive Decentralized Monitoring in Software-Defined Networks

“…For each experiment, only one of the two services is emulated. In the case of content distribution, each client generates content requests following a pattern derived from the one used in [37]. The content size is scaled down in accordance to the reduced link bandwidth.…”

Decentralized monitoring for large-scale Software-Defined Networks

“…The placement of LMs is provided as an input and used to compute the relevant hop-count and corresponding latencies between pairs of LMs. After the startup, each client starts generating content requests following a pattern derived from the one used in [34]. The content size has been scaled down in accordance to the reduced link bandwidth.…”

Low complexity content replication through clustering in Content-Delivery Networks

“…In reality, content popularity changes significantly over time [22][23] [24] and the design of a content replication scheme that updates caches following closely those changes is a challenging management task, since replication is a time-consuming process.…”

“…Content popularity is changing over time [22][23] [24], and based on the request rates observed within a time window of duration T obs , the manager/central entity of the network may detect substantial variations in popularity, say at time T 0 = 0. We model this modification of request vectors through a popularity alteration factor a in the sense that the ranking of the items within the Zipf popularity distribution at each node is altered by a factor a; i.e., a · N items have a different ranking at T 0 than the most recent one that was used to derive the current replication decision.…”