Algorithms for Server Rebuild - Scalable Continuous Media Streaming Systems - page 259

800

600

400

200

0

0 0.2 0.4 0.6 0.8 1

System Utilization

Reload from back-up
Baseline rebuild
Distributed rebuild
Mixed distributed-baseline rebuild

Figure 14.7 Data rebuild rate versus system utilization

Figure 14.8 shows the time required to completely rebuild a server's worth of data (200GB).
Interestingly, for system utilization

75 performances of the baseline rebuild algorithm
and themixed distributed baseline algorithmconverge. This is because at high systemutilization
the remaining active servers become the bottleneck, and so it is better to switch to the baseline
rebuild algorithm which consumes less transfer capacities of the active servers than distributed
rebuild. In this case the mixed distributed baseline algorithm simply allocates all transfer
capacity to baseline rebuild.
Finally, we plot in Figure 14.9 the rebuild time versus the number of servers in the system
under a systemutilization of

ρ ≥ 0 .

5. The key observation is that the baseline rebuild algorithm
is not scalable - the rebuild time increases with the number of servers in the system. This is
because in baseline rebuild the maximum rebuild rate is limited by the transfer capacity of the
spare server. Thus, as the number of servers increases, so will the number of data/redundant
units that need to be sent to the spare server to rebuild an unavailable data unit, thereby resulting
in longer rebuild time.
By contrast, the rebuild times of distributed rebuild and mixed distributed baseline rebuild
do not increase significantly with increases in the number of servers and so are much more

ρ = 0 .