Information Technology Reference
In-Depth Information
the response time for PSO, ACO, and GA, when the number of jobs was increased from
1,000 to 10,000, varied between (79-668), (79-668), and (367-962) minutes. Finally,
when the LTIR policy is used, the response time rose from (79- 668), (84-673), and
(409-996) minutes, when the number of jobs was increased from 1,000 to 10,000, and
for PSO, ACO and GA, respectively.
As can be seen, the response times for ACO and PSO are close when FLTF and LTIR
are used at the broker level. The reason is because both algorithms reduce the number
of queries to the hosts when LTIR is used. This is because when ACO and PSO find
an idle host, they not make any further move, and due to the fact that LTIR explores
all datacenters (in a circular order for each VM to be allocated), it has more chance of
finding an underloaded hosts where to allocate the VMs. However, if the user requests
the execution of a larger number of VMs, the latencies of datacenters will have more
influence in the response time when LTIR is used instead of FLTF.
Finally, the gains of PSO and ACO with respect to GA, when LLTF is used at the
broker level varied between 15% and 62%. When FLTF was used, gains varied between
30% and 78%. Lastly, when LTIR was used, gains varied between 32% and 80%. As
can be seen, the greatest gains were obtained when LTIR was used at the broker level.
The is because, since GA sent a greater number of network messages to the hosts than
PSO and ACO, the inter-datacenter latencies had more influence on the response time.
6
Conclusions
One popular kind of scientific experiments are PSEs, which involve running many CPU-
intensive independent jobs. These jobs must be e
ciently processed -i.e., scheduled-
in the di
erent computing resources of a distributed environment such as the ones pro-
vided by Cloud. The growing popularity of Cloud environments has increased the at-
tention in the research of resource allocation mechanisms across datacenters. Federated
Clouds potentially provide plenty of resources to users, specially when the number of
VMs required by a user exceeds the maximum that can be provided by a single provider
or datacenter. Then, job
ff
VM scheduling plays a fundamental role.
Recently, SI-inspired algorithms have received increasing attention in the Cloud re-
search community for dealing with VM and job scheduling. In this work, we described
two schedulers -based on ACO and PSO- for the e
/
cient allocation of VMs in a dat-
acenter combined with three strategies -LLTF, FLTF and LTIR- that consider network
information for selecting datacenters. Simulated experiments performed with CloudSim
and real PSE job data suggest that our PSO and ACO schedulers provide better response
times to the user than GA. In addition, when PSO, ACO and GA are combined with
LLTF, the response time is the lowest for all of them w.r.t. FLTF and LTIR, being LTIR
the most influential on the response time.
We are extending this work in several directions. We will explore the ideas exposed in
this paper in the context of other bio-inspired techniques such as Artificial Bee Colony
(ABC), which is also extensively used to solve combinatorial optimization problems.
Another issue which deserves attention is to consider other Cloud scenarios [15] with
heterogeneous physical resources belonging to di
erent Cloud providers.
Due to multi-tenancy, in Clouds it is necessary to provide distributed scheduling
mechanisms for allocating resources to a number of independent users' VMs
ff
/
jobs along
