Deploying a Multi-interface RESTful Application in the Cloud - Data Management in Cloud, Grid and P2P Systems - page 83

10000

No Affinity
Affinity

8000

6000

4000

2000

0

0 5 10 15 20

Time (min)

Fig. 5 . Impact of session anity on performance

100 40000

75 30000

50 20000

25 10000

0 0

0 10 20 30 40 0 10 20 30 40

Time (min)
(a) Average CPU utilization

Time (min)
(b) Sum of requests per 5 minute period

Fig. 6 . Experimental results of scaling operation using the small instance type
Autoscaling Tests. The initial 10Green experimental deployments used the
micro instance because of its low cost. Utilizing many available servers, even
with modest memory and processing speed, is well suited in particular for the
tile mapping system where many tiles may be requested by the same client. For
these tests a variable number of client instances were used (5 initially, 5 added
every five minutes to a maximum of 25), caching was disabled, and the scaling
increment was two new server instances.
The first experiment, using a micro instance type, resulted in an almost instant
failure of the application. Even though the framework noticed the increase in load
and spawned additional instances, the limited I/O performance of the micro
instance was a major bottleneck, causing the application to be unavailable.
The experiment with the small instance, on the other hand, yielded a deploy-
ment that scaled successfully. In particular, it demonstrated an ability to manage
significant spikes in usage, with very limited periods of high latency. Some results
of this experiment are summarized in Fig. 6. As the rate of requests increased
(Fig. 6b), the average CPU utilization (Fig. 6a) crosses the 50% threshold and
causes two new instances of the application server to be spawned, which in turn
causes a drastic decrease in the utilization. Another CPU usage peak occurs
as the number of requests climbs even further, but is quickly reduced by the
instantiation of another pair of server instances.