Boosting Streaming Video Delivery with WiseReplica - Transactions on Large-Scale-Data-and Knowledge-Centered Systems XX - page 36

As the popularity of a video varies, the number of replicas, or peers serving that
video, must be adapted accordingly. Generally speaking, the faster and more
precise the replication scheme reacts to changes on videos demand, the better is
the resource allocation and content availability.
Considering average bitrate as target QoS metric, we make a case for a
SLA-driven replication scheme named WiseReplica that allows us to meet users'
expectations in peer-assisted VoD system properly. We assume the system must
enforce the right average bitrate for each video through SLA contracts. Our
ultimate goal is two-fold: (i) to prevent SLA violations and (ii) to reduce the
number of video replicas. To perform ecient Internet video replication, Wis-
eReplica relies on a novel, accurate machine-learned ranking of Internet videos.
To rank video in order of demand, our prediction model encompasses multiple
measurements of Internet video activity in peer-assisted VoD system, including
active viewers, video duration, average serving time, and mean time between
requests view. The use of this prediction model in WiseReplica provides the
ability to adapt the replication degree of videos dynamically according to their
encoding settings and popularity, reducing storage usage and enhancing network
provision. We make two main contributions:
Investigate how predictable is a ranking of Internet videos. We design
a learning model to capture the dynamic behaviour of streaming video demand.
The model makes predictions based on lightweight measurements of the request
arrival process. Using a novel machine-learned ranking, we predict demand of a
video accurately. Thus, the higher the rank position, the higher the demand for
fresh replicas. According to the video ranking position, VoD services operators
can define and evaluate different replication policies. For instance, top-ranked
Internet videos may be twice as much replicated as those ranked in the second
position. This intuitive model allows us to decouple streaming demand from
replication policy. Our model is flexible and can learn from different sources and
big amounts of data, providing a robust framework for controlling VoD resource
allocation. Simulations using YouTube traces, with non-stationary behaviours,
suggest that our model is very accurate in predicting the ranking of Internet
videos. Since our ranking of videos is based on random forests, a parallelizable,
state-of-the-art machine learning method, it fits runtime requirements of large
VoD systems.
Enforce average bitrate through SLA-based video replication. Based
on our machine-learned ranking of Internet video, we designed and evaluated
WiseReplica, an easy-to-deploy, SLA-based replication scheme that meets users'
expectations for VoD services. WiseReplica is fully compliant with peer-assisted
VoD systems in hybrid CDN platforms. It operates adaptive replication over sets
of devices located close to each other in edge networks, namely storage domains .
WiseReplica functioning per storage domain is straightforward. Gradually, it
verifies the rank position of a video whenever a new local request arrives, and
adapts the replication degree accordingly. Using a collaborative caching, video
replicas are either pre-fetched or removed randomly. We show through simula-
tions using YouTube traces that WiseReplica outperforms a non-collaborative