Boosting Streaming Video Delivery with WiseReplica - Transactions on Large-Scale-Data-and Knowledge-Centered Systems XX - page 45

with SLA violations and meet customers' expectations, we evaluate four quite
simple policies, namely uniform, linear, quadratic, and exponential. They are
respectively defined as follows:

2 ,and

B r , where

B, Br, Br B is a constant that

represents the target number of replicas, and r ∈{ 1 , 2 , 3 } the rank positions.

We report on creation policies' performances in Sect. 6 .
Our findings show that this approach produces a good balance between
resource usage and consumers' satisfaction. It is important to note, however,
WideReplica does not cover video durability, neither does fault-tolerant mecha-
nisms (e.g. failure detection/recovery procedures). Rather, our goal is to improve
VoD availability, boosting network provision, meeting consumers' expectations
on VoD services, and reducing storage usage as much as possible. To this end,
WiseReplica combines lightweight measurements, accurate predictions of Inter-
net videos ranking, and replication policies enforcement in a particularly novel,
flexible way. In peer-assisted VoD systems, it can easily interoperate with de
facto approaches, including HTTP adaptive streaming technique and swarming
protocols, such as BitTorrent.

5 Simulation Methodology

We simulate a peer-assisted VoD system based on a hybrid CDN design
called Caju [ 30 ]. We evaluate WiseReplica using YouTube traces. We compare
WiseReplica performance with other two adaptive replication schemes, namely
non-collaborative caching and Oracle-like collaborative caching. The aim of our
simulations is to study in details the variability of demand and resource allo-
cation of VoD services on edge networks, and the performance of replication
schemes in enforcing expected Internet video availability.

5.1 Workload from YouTube Traces and SLA Definition

The workload and SLA definitions are at the core of our evaluation. We define a
workload that captures the main features of VoD services using YouTube traces,
and a SLA contract that meets users' expectations.
In the workload definition, we are particularly interested in reproduce a real-
ist request arrival process, placing the emphasis on popularity growth and video
encodings. Thus, we use YouTube traces, presented in Subsect. 2.2 . Before inte-
grating YouTube traces to our workload, we first preprocessed their YouTube
datasets to remove inconsistent measurements, such as videos with no views.
Basically, we got rid of videos with small number of total views (those smaller
than the first quartile) and videos with few daily measurements (those smaller
than the third quartile). That allowed us to pick off 20 % most representative
YouTube growth patterns, accounting for 21827 distinct curves. Then, we ran-
domly selected, with a uniform distribution, curves from this preprocessed data
to be assigned to videos of our workload. Similarly, we assigned high quality
YoutTube video encodings to our workload videos, based on advanced settings
depicted in Table 1 . To summarize, Table 2 lists default values for workload