Methods to enhance content distribution for very large scale online communities

摘要

The Internet has experienced an exponential growth in the last years, and its number of users far from decay keeps on growing. Popular Web 2.0 services such as Facebook, YouTube or Twitter among others sum millions of users and employ vast infrastructures deployed worldwide. The size of these infrastructures is getting huge in order to support such a massive number of users. This increment of the infrastructure size has brought new problems regarding scalability, power consumption, cooling, hardware lifetime, underutilization, investment recovery, etc. Owning this kind of infrastructures is not always affordable nor convenient. This could be a major handicap for starting projects with a humble budget whose success is based on reaching a large audience. However, current technologies might permit to deploy vast infrastructures reducing their cost. We refer to peer-to-peer networks and cloud computing. Peer-to-peer systems permit users to yield their own resources to distributed infrastructures. These systems have demonstrated to be a valuable choice capable of distributing vast amounts of data to large audiences with a minimal starting infrastructure. Nevertheless, aspects such as content availability cannot be controlled in these systems, whereas classic server infrastructures can improve this aspect. In the recent time, the cloud has been revealed as a promising paradigm for hosting horizontally scalable Web systems. The cloud offers elastic capabilities that permit to save costs by adapting the number of resources to the incoming demand. Additionally, the cloud makes accessible a vast amount of resources that may be employed on peak workloads. However, how to determine the amount of resources to use remains a challenge. In this thesis, we describe a hierarchical architecture that combines both: peer-to-peer and elastic server infrastructures in order to enhance content distribution. The peer-topeer infrastructure brings a scalable solution that reduces the workload in the servers, while the server infrastructure assures availability and reduces costs varying its size when necessary. We propose a distributed collaborative caching infrastructure that employs a clusterbased locality-aware self-organizing P2P system. This system, leverages collaborative data classification in order to improve content locality. Our evaluation demonstrates that incrementing data locality permits to improve data search while reducing traffic. We explore the utilization of elastic server infrastructures addressing three issues: system sizing, data grouping and content distribution. We propose novel multi-model techniques for hierarchical workload prediction. These predictions are employed to determine the system size and request distribution policies. Additionally, we propose novel techniques for adaptive control that permit to identify inaccurate models and redefine them. Our evaluation using traces extracted from real systems indicate that the utilization of a hierarchy of multiple models increases prediction accuracy. This hierarchy in conjunction with our adaptive control techniques increments the accuracy during unexpected workload variations. Finally, we demonstrate that locality-aware request distribution policies can take advantage of prediction models to adequate content distribution independently of the system size.