Database Reference
In-Depth Information
Consistent Read
Eventual Consistent Read
14.0
12.0
99.9 percentile
10.0
95.0 percentile
8.0
6.0
4.0
Average
2.0
0.0
0
500
1000
1500
2000
2500
Number of Emulated Clients (Threads)
Fig. 5.4
The average, 95 percentile, and 99:9 percentile response time of reads at various levels
of load
Response Time and Throughput
As advised in Amazon SimpleDB FAQs,
1
the benefits of eventual consistent read
can be summarized as minimizing response time and maximizing throughput.
To verify this advice, the difference in response time, throughput, and availability of
the two consistency options is investigated, as the load is increased. Figure
5.4
shows
the average, 95 percentile, and 99:9 percentile response time of eventual consistent
reads and consistent reads at various levels of load. The result is obtained from
the case of 99% read ratio and all failed requests are excluded. The result shows
no visible difference in average response time. However, consistent read slightly
outperforms eventual consistent read in 95 percentile and 99:9 percentile response
time.
Figures
5.5
and
5.6
show the average response time of reads and writes at various
read/write ratios, plotted against the number of emulated end-users. A conclusion
could be drawn that changing the level of replication intensity has a negligible
impact on the read and write response times. Intuitively, it would be surprised that
eventual consistent read does not outperform the consistent read as expected, but it
is still reasonable if the possible implementation ideas are taken into consideration.
Figure
5.7
shows the absolute throughput, the average number of processed requests
per second. Whiskers are plotted surrounding each average with the corresponding
minimum and maximum throughput. Similar to the response time, consistent read
results slightly outperforms that of eventual consistent read, though the difference is
not significant. Figure
5.8
shows the throughput as a percentage of what is possible
