5.3. Case study: hybrid cloud provisioning strategy
In this section, a more complete experiment that also captured the networking behavior (latencies)
between clouds is presented. This experiment showed that the adoption of a hybrid public/private
Cloud computing environments could improve the productivity of a company. With this model,
companies can dynamically expand their system capacity by leasing resources from public clouds
at a reasonable cost.
The simulation scenario models a network of a private and a public cloud (Amazon EC2
cloud). The public and the private clouds were modeled to have two distinct data centers. A
CloudCoordinator in the private data center received the user’s applications and processed (queue,
execute) them on an FCFS basis. To evaluate the effectiveness of a hybrid cloud in speeding up
tasks execution, two test scenarios were simulated: in the first scenario, all the workload was
processed locally within the private cloud. In the second scenario, the workload (tasks) could be
migrated to public clouds in case private cloud resources (hosts, VMs) were busy or unavailable.
In other words, the second scenario simulated a Cloud-Burst by integrating the/a local private
cloud with public cloud for handing peak in service demands. Before a task could be submitted to
a public cloud (Amazon EC2), the first requirement was to load and instantiate the VM images at
the destination. The number of images instantiated in the public cloud was varied from 10 to 100%
of the number of hosts available in the private cloud. Task units were allocated to the VMs in the
space-shared mode. Every time a task finished, the freed VM was allocated to the next waiting
task. Once the waiting queue ran out of tasks or once all tasks had been processed, all the VMs
in the public cloud were destroyed by the CloudCoordinator.
The private cloud hosted approximately 100 machines. Each machine had 2GB of RAM, 10 TB
of storage, and one CPU run 1000 MIPS. The VMs created in the public cloud were based on
an Amazon’s small instance (1.7GB of memory, 1 virtual core, and 160GB of instance storage).
We considered in this evaluation that the virtual core of a small instance has the same processing
power as the local machine.
The workload sent to the private cloud was composed of 10 000 tasks. Each task required
between 20 and 22min of processor time. The distributions for processing time were randomly
generated based on the normal distribution. Each of the 10 000 tasks was submitted at the same
time to the private cloud.