Boosting Streaming Video Delivery with WiseReplica - Transactions on Large-Scale-Data-and Knowledge-Centered Systems XX - page 42

WiseReplica, in order to identify YouTube ranking patterns. Once the learning
phase has been accomplished, WiseReplica can use its learning module in a pre-
dicting phase, as indicated in the left-hand side of Fig. 1 . In this phase, inputs
come for measurements of the request arrival process of workload 2, that permit
accurately ranking Internet videos in order of hotness and instrumenting replica-
tion accordingly inside storage domains. We highlight WiseReplica functioning,
including storage domains, in the next section.
4 Boosting VoD Delivery: The WiseReplica Approach
In this section, we describe WiseReplica replication scheme. First, we highlight
how WiseReplica operates in edge networks, by in introducing the concept of
storage domains. Then, we explain its replication strategy based on predictions
of ranking of video demand.
4.1 Distributing VoD with Storage Domains
We assume that WiseReplica operates in peer-assisted VoD systems deployed on
hybrid CDN platforms. We consider the hybrid CDN design called Caju, that is
detailed in our previous work [ 30 ] as our target platform. It is based on sets of
devices located close to customers, named storage domain. A storage domain is a
logical entity that combines resources from both datacenters and edge networks
in the last mile of the content delivery chain. As Fig. 2 shows, devices in a storage
domain can play either a coordinator or peer role.
Coordinator is a server or a small-sized cluster of servers deployed in the
nearby datacenter. We assume that the coordinator performs scheduling of video
requests for the local storage domains. Therefore, it runs the main instance of
WiseReplica, and keeps information about resources consumption. Its main goal
is to maintain the right number of replicas per video in the local peers, by pre-
fetching or deleting sources. Instead of always contacting the content providers,
coordinators might interoperate in logically centralized way to fetch videos that
have been vanished from a storage domain. They store the most recent videos
in their own cache for replication purposes. Whenever a new replica is neces-
sary, the coordinator pushes it to a randomly, uniformly selected peer. Similarly,
coordinators send video deletion requests to local peers.
Peers is a set of devices located close to each other through which customers get
network access, e.g. home gateways connected to the same digital subscriber line
access multiplexer (DSLAM). These devices actually deliver videos to customers
in a storage domain, being the main source of storage and network resources.
They execute scheduling and replication commands sent by the storage domain's
coordinator. Each peer contributes with a percentage of storage and network
resources to the system, as in a collaborative caching. In the local cache is applied
the LRU policy for videos replacement.
This model is specially interesting for the problem of videos delivery as
it takes advantage of nodes geographical position [ 6 ]. It provides two main