CloudDB AutoAdmin - Large Scale and Big Data: Processing and Management - page 377

workload gradually increases in steps of 20 concurrent users every 600 seconds so
that each experiment ends with a workload of 300 concurrent users. Each experiment
deploys 6 replicas in 3 regions where each region hosts two replicas: the first replica
is an active replica that is used from the start of the experiment for serving the data-
base requests of the application, while the second replica is a hot backup that is not
used for serving the application requests at the beginning of the experiment but can
be added by the action module, as necessary, when triggered by the control module.
Finally, in addition to the two sets of experiments, we conducted two experiments
without our adaptive replication controller to measure the end-to-end throughputs
and replication delays of 3 (the minimum number of running replicas) and 6 (the
maximum number of running replicas) slaves in order to measure the baselines of
our comparison.
11.6.2.1 End-to-End Throughput
Table 11.1 presents the end-to-end throughput results for our set of experiments with
different configuration parameters. The baseline experiments represent the minimum
and maximum end-to-end throughput results with 22.33 and 38.96 operations per
second, respectively. They also represent the minimum and maximum baseline for
the running time of all database replicas with 9000 (3 running replicas, with 3000
seconds running time of each replica from the beginning to the end of the experi-
ment) and 18,000 (6 running replicas, with 3000 seconds running time of each rep-
lica) seconds, respectively. The end-to-end throughput of the other experiments fall
between the two baselines based on the variance on the monitor interval ( intvl mon ) and
the tolerance of replication delay ( delay tolerance ). Each experiment starts with 3 active
replicas after which the number of replicas gradually increases during the experi-
ments based on the configurations of the monitor interval and the SLA of replication
delay parameters until it finally ends with six replicas. Therefore, the total running
time of the database replicas for the different experiments fall within the range
between 9000 and 18,000 seconds. Similarly, the end-to-end throughput delivered
by the adaptive replication controller for the different experiments fall within the
end-to-end throughput range produced by the two baseline experiments of 22.33 and
38.96 operations per second. However, it is worth noting that the end-to-end through-
put can be still affected by a lot of performance variations in the cloud environment
such as hardware performance variation, network variation, and warm up time of
the database replicas. In general, the relationship between the running time of all
slaves and end-to-end throughput is not straightforward. Intuitively, a longer monitor
interval or a longer tolerance of replication delay usually postpones the addition of
new replicas and consequently reduces the end-to-end throughput. The results show
that the tolerance of the replication delay parameter ( delay tolerance ) is more sensitive
than the monitor interval parameter ( intvl mon ). For example, having the values of the
tolerance of the replication delay equal to 4000 and 1000 result in longer running
times of the database replicas than having the values equal to 2000 and 500. On the
other side, the increase of running time of all replicas shows a linear trend along with
the increase of the end-to-end throughput. However, a general conclusion might not
be easy to draw because the trend is likely affected by the workload characteristics.