Information Technology Reference
In-Depth Information
to a host. To create larger selection probability, a larger central angle was assigned to
a sector associated with a host that has larger amount of available resources. In the
proposed VM mapping policy, multi-dimensional resource usage (e.g., CPU and
memory) was considered. However, other constraints, such as electricity saving and
high availability of a virtual system, were not considered.
Xu et al. [9] and Garces et al. [10] implemented multi-objective optimization
mechanisms for VM placement and migration. Their common approach applies a
genetic algorithm to solve a multi-objective optimization problem. Our approach does
not use a genetic algorithm. As mentioned in Section 4, we introduced equivalent sets
of hosts to avoid extremely large numbers of possible placement combinations.
Tsakalozos et al. [11] proposed an approach similar to ours; i.e., identifying
potentially compatible groups of physical servers to reduce the search space.
Moreover, a few constraints such as power saving and minimizing network traffic by
co-deploying a set of VMs on the same physical server were considered. In this
research, however, physical servers were classified into groups based solely on VM
migration ability because the goal was load balancing through migration. In contrast,
our mechanism generates equivalent sets of physical servers (hosts) for each
constraint or policy. Note that the VM migration ability of a host can be added as a
constraint by implementing a plug-in module. Subsequently, our mechanism places a
VM on a host that was evaluated as having the lowest comprehensive cost because
our goal is the reduction of operational cost and not load balancing.
8
Conclusion
We have designed and implemented a multi-objective optimization mechanism for
VM placement. The proposed mechanism is flexible and allows data center operators
to add their own desired optimization objectives or evaluate specific policies by
implementing plug-in modules.
The unique value of our system is that a constraint is translated into an estimated
cost (real or assigned). Each plug-in module evaluates the additional cost that would
accrue if a VM is placed on a host. Subsequently, the Comprehensive-Evaluation
Module gathers the results and calculates the total cost considering the weight of each
policy. The Total optimization controller conducts a neighborhood search to find the
lowest cost VM placement, and ultimately, it enforces the optimum VM placement.
A practical calculation time to find an optimum VM placement was realized by
introducing the equivalent sets of hosts. The simulation results showed that both
availability and repulsion could be satisfied with a cost saving of 8.4%-27.3% for
electricity.
Acknowledgments.
We would especially like to express our heart-felt gratitude to
Messrs. Toru Sakiyama, Noriyuki Murata, and Makoto Nasuno who are systems
software engineers at ASCADE Inc.
