On the design of efficient caching systems

Content distribution is currently the prevalent Internet use case, accounting for the majority of global Internet traffic and growing exponentially. There is general consensus that the most effective method to deal with the large amount of content demand is through the deployment of massively distributed caching infrastructures as the means to localise content delivery traffic. Solutions based on caching have been already widely deployed through Content Delivery Networks. Ubiquitous caching is also a fundamental aspect of the emerging Information-Centric Networking paradigm which aims to rethink the current Internet architecture for long term evolution. Distributed content caching systems are expected to grow substantially in the future, in terms of both footprint and traffic carried and, as such, will become substantially more complex and costly. This thesis addresses the problem of designing scalable and cost-effective distributed caching systems that will be able to efficiently support the expected massive growth of content traffic and makes three distinct contributions. First, it produces an extensive theoretical characterisation of sharding, which is a widely used technique to allocate data items to resources of a distributed system according to a hash function. Based on the findings unveiled by this analysis, two systems are designed contributing to the abovementioned objective. The first is a framework and related algorithms for enabling efficient load-balanced content caching. This solution provides qualitative advantages over previously proposed solutions, such as ease of modelling and availability of knobs to fine-tune performance, as well as quantitative advantages, such as 2x increase in cache hit ratio and 19-33% reduction in load imbalance while maintaining comparable latency to other approaches. The second is the design and implementation of a caching node enabling 20 Gbps speeds based on inexpensive commodity hardware. We believe these contributions advance significantly the state of the art in distributed caching systems