Boosting Streaming Video Delivery with WiseReplica - Transactions on Large-Scale-Data-and Knowledge-Centered Systems XX - page 53

Table 3. Replication policies.
Load Resiliency. A good replication strategy must cope with changes on the
system load. We vary the global load of the system by changing the mean video
size, described in Subsect. 5.1 . Assuming the three mean video sizes, namely
20 MB, 30 MB and 40 MB, caching had 1814, 3864, and 7049 violations respec-
tively, while WiseReplica had only 6, 77, and 106. Figure 9 compares the number
of violations using WiseReplica and a non-collaborative caching. As the load of
the system increases, concurrency in bitrate allocation also increases, causing
more violations. WiseReplica outperforms caching mostly because it predicts
and prevents useless replication. Therefore, we set to the highest evaluated sys-
tem load, 40 MB, as the default mean video size workload setting.
Benefits of Prediction on Storage Usage. We aim to adapt the number
of replicas to the number of views of a video, especially for the most popular
ones. Figure 10 plots the maximum number of replicas for the 1 % most popular
videos. Using caching, the maximum number of replicas is high, ranging from 816
to 1367. The Oracle-like assumption allows to decrease significantly the lower
and upper limits, to 10 and 190. WiseReplica also reduces the maximum replica
range, which is from 19 to 160. More interestingly, the shape of the replication
Fig. 9. Mean video size. Higher loads increase the concurrence in network resources,
as a result, more violations.