Boosting Streaming Video Delivery with WiseReplica - Transactions on Large-Scale-Data-and Knowledge-Centered Systems XX - page 39

adaptive replication scheme should offer content replica maintenance to handle
popularity growth properly.
Non-collaborative caching remains the simplest approach to provide adaptive
replication of web content [ 20 ]. They adapt the replication degree to the con-
tent popularity using cache replacement policies, and assuming fair-sharing as
a key scheduling strategy. But, Internet videos' workloads on peer-assisted VoD
systems bring major obstacles to non-collaborative caching, e.g. the resource
imbalance in peers for replicas, and a growing need for high bitrate provision for
meeting consumers' expectations. Therefore, relying just on cache replacement
policies and fair-sharing scheduling can undermine the performance of the whole
system.
Recent studies have sought an optimal solution to this problem. For instance,
Chang and Pan [ 10 ] propose a modelling framework towards optimal caching
strategies, including collaborative caching. They confirm that this problem is
NP-hard, and only suboptimal solutions can be found.
2.4 Challenges
In order to meet increasing consumers' expectations on Internet videos, a good
peer-assisted VoD system must overcome the following challenges:
1. It must cope with dramatic, unexpected variations in videos popularity.
2. It must avoid waste of resources, and reduce as much as possible storage and
network usage on peers of edge networks.
3. It must prevent rebuffering of VoD streaming through a self-adaptive, easy-
to-deploy technique.
Our simulations suggest that meeting consumers' expectations in terms of
average bitrate is a dicult task, specially under heavy load. State-of-the-art
approaches fail to handle these challenges mostly because they are not able
(i) to capture VoD demand, and (ii) to define a metric to measure consumers'
expectations. WiseReplica copes with these issues by inferring users' expecta-
tions for videos and predicting the amount of resources to fulfil the demand in a
self-adaptive way. Our findings show that this approach produces a good balance
between resource allocation and users' satisfaction.
3 A Machine-Learned Ranking of Internet Videos
We designed a prediction model for ranking Internet videos in order of demand.
In this work, video demand involves both popularity and QoS requirements.
Our main goal is to provide an intuitive, accurate method to capture requesting
behaviours of streaming videos. In this section, we highlight the foundations of
our statistical learning approach. First, we present a brief overview of statistical
learning. Then we explain the model, describing our learning-to-rank problem.
Finally, we describe our implementation and we present a framework for ranking
predictions.