Database Reference
In-Depth Information
Fig. 7.
Accuracy with different sample
sizes.
Fig. 8.
Relative importance to ranking
of the 10 model's inputs.
size seem to be useless to our model. It is likely that network availability is a
particularly good measurement, making these eight inputs rather redundant. For
simplicity, we include all inputs in the rest of the work. This is harmless for the
model's accuracy.
6.3 Evaluating Replication Strategies in Peer-Assisted VoD
Systems
In this subsection we analyse the replication strategy used in WiseReplica. First,
we evaluate four simple replication policies. Then, we compare WiseReplica with
a non-collaborative caching and Oracle-like benchmark replication approach,
both described in Subsect.
5.3
. We evaluate their capacity to meet consumers'
expectation by observing the number of violations. In addition, we compare their
resource allocation performance regarding network and storage usage.
Enforcing Simple Replication Policies on Ranked VoD.
For the three
highest rank position, WiseReplica enforces a replica creation policy, described
in Subsect.
4.2
. It defines the replication degree growth factor. Considering the
smallest evaluated system load (with mean video size of 20 MB), we analyse
four simple creation policies, namely uniform, linear, quadratic, and exponen-
tial. Table
3
shows the number of violations by varying
from 2 to 6. Overall,
creation policies that take into account the rank positions, i.e. linear, quadratic,
and exponential, performed better. Results show that there is relatively small
difference for
B
3, suggesting that our ranking model reacts promptly to modi-
fications on network availability, preventing over-replication. However, for
B ≥
5,
it appears that replication increases the network load system load, causing few
more violations. We selected the linear policy with
B ≥
= 4 that seems to be the
most resilient towards proper resource allocation, providing a fair replication
degree growth factor.
B
Search WWH ::
Custom Search