CloudDB AutoAdmin - Large Scale and Big Data: Processing and Management - page 372

caused by the performance variation of instances rather than the configuration of
locations. The first slave from the same zone runs on top of a physical machine
with an Intel Xeon CPU E5430 2.66GHz, while another first slave from the differ-
ent zone is deployed in a physical machine powered by an Intel Xeon CPU E5507
2.27GHz. Because of the performance differences between physical CPUs, the slave
from same zone performs better than the one from different zone. Previous research
indicated that the coefficient of variation of CPU of small instances is 21% [18].
Therefore, it is a good strategy to validate instance performance before deploying
applications into the cloud, as poor-performing instances are launched randomly and
can largely affect application performance.
11.5.4 r ePliCation D elay e XPeriments
Figures 11.6 and 11.7 show the trends of the average relative replication delay for up
to 4 and 11 slaves with mixed configurations of three locations and two read/write
ratios. The results of both figures imply that the configurations of the geographical
locations have a lower impact on the replication delay than that of the workload
characteristics. The trends of the average relative replication delay respond to an
increasing workload and an increasing number of database replicas. For most cases,
with the number of database replicas being kept constant, the average relative rep-
lication delay surges along with an increasing workload, which leads to more read
and write operations sent to the slaves and the master database, respectively. It turns
out that the increasing number of read operations result in a higher resource demand
on every slave while the increasing write operations on the master database leads to,
indirectly, increasing resource demand on slaves as more writesets are propagated
to be committed on slaves. The two increasing demands push resource contention
higher, resulting in the delay of committing writesets, which subsequently results in
higher replication delay. Similarly, the average relative replication delay decreases
along with an increasing number of database replicas as the addition of a new slave
leads to a reduction in the resource contention and subsequent decrease in replica-
t ion delay.
As previously mentioned, the configuration of the geographic location of the
slaves play a less significant role in affecting replication delay, in comparison to
the changes of the workload characteristics. We measured the half round-trip time
between the master in us - west -1 a and the slave that uses different configurations of
geographic locations by running the ping command every second for a 20-minute
period. The results suggest an average of 16, 21, and 173 milliseconds half round-
trip time for the same zone (Figures 11.6a and 11.7a), different zones (Figures 11.6b
and 11.7b), and different regions (Figures 11.6c and 11.7c), respectively. However, the
trends of the average relative replication delay can usually go up from two to four
orders of magnitude (Figure 11.6), or one to three orders of magnitude (Figure 11.7).
Therefore, it could be suggested that geographic replication would be applicable in
the cloud as long as workload characteristics can be well managed (e.g., having a
smart load balancer that is able to balance the operations based on estimated pro-
cessing time).