Boosting Streaming Video Delivery with WiseReplica - Transactions on Large-Scale-Data-and Knowledge-Centered Systems XX - page 44

Internet videos delivery is a resource-hungry service, we must also adapt the
network provision and storage usage as we aim to prevent violations. To this
end, we propose WiseReplica, an adaptive replication scheme for peer-assisted
VoD systems based on storage domains.
WiseReplica maintains replicas inside a storage domain. Running on the
coordinator, it adapts the replication degree of Internet videos of a storage
domain according to a machine-learned ranking. Our scheme follows a three-
part procedure:

Collect Information from the Request Arrival Process. For each video
request, WiseReplica collects 10 lightweight measurements. The goal is to gather
comprehensive information for measuring the video demand and accurately pre-
dicting the raking. As described in Sect. 3 , they are video size, network availabil-
ity, network usage (load), current number of viewers and replicas, inter-arrival
time between requests (delta), aggregate number of views, mean of time between
requests (mtbr), life time, and average bitrate. We compute averages and means
from the last up to five requests. It is important to notice that all these mea-
surements can be easily collected in the storage domain's coordinator.

Rank Internet Videos in Order of Demand. Based on the measurements
of the request arrival process, we use the learning model described in Sect. 3
for predicting the rank position of Internet videos demand. We can predict the
video rank for each view from the second request. The ranking comprises infor-
mation about demand and QoS requirements. Predictions are quite essential
for enhancing VoD delivery. Since our learning model make predictions on a
request basis, WiseReplica can react to the video demand as promptly as the
rank position evolves. Indeed, ranking is an intuitive way to capture the demand
of videos in peer-assisted VoD systems. The higher is the demand rank position
of an Internet video, the higher is the demand for it. There are four positions
on our machine-learned ranking: non-popular, popular, very popular, and viral .
WiseReplica has a straightforward strategy to perform replication according to
hotness rank positions. Videos that fall into the lowest rank position can have
their replication degree reduced, otherwise they need more replicas. Thus, the
maintenance of replication degree of Internet video, including video creations
and deletions in peers, relies on replication policies.

Enforce Replication Policy Accordingly and in Time. The goal of repli-
cation policies is two-fold: first (i) ensure consumers' expectations in time and
(ii) reduce the total number of replicas as much as possible. For that, WiseReplica
must adapt replication of videos according to the forecasts of their rank positions.
Our replication scheme enforces two types of replica maintenance policies: dele-
tion and creation policy. In this work, we enforce a single video deletion policy.
Whenever the coordinator receives a request to a video in the non-popular rank,
the deletion policy says that one replica is deleted until the minimum replication
degree

m

is reached. Similarly, our scheme periodically runs a maintenance pro-
cedure (e.g. each five minutes) to smoothly enforce the deletion policy for inactive
videos. This allows WiseReplica to reduce the total number of replicas. To cope