Information Technology Reference
In-Depth Information
them. Secondly, at the infrastructure level, by using a VM scheduler, the VMs are allo-
cated on the available hosts belonging to the previously selected datacenters. Lastly, at
the VM level, by using job scheduling techniques, jobs are assigned for execution into
allocated virtual resources. However, scheduling is in general an NP-Complete [21]
problem and therefore it is not trivial from an algorithmic standpoint. Besides, in this
context, the necessity of scheduling algorithms spans the three levels.
In the last ten years, Swarm Intelligence (SI) has received increasing attention among
researchers. SI refers to the collective behavior that emerges from a swarm of so-
cial insects [9]. Inspired by these capabilities, researchers have proposed algorithms
or theories for combinatorial optimization problems, where the most popular SI-based
strategies are Ant Colony Optimization (ACO) and Particle Swarm Optimization (PSO).
Moreover, job scheduling in Clouds is also a combinatorial optimization problem, and
schedulers in this line that exploit SI have been proposed.
Existing e
orts which address SI have not being studied in the context of federated
Clouds. In this paper, unlike previous works of our own [16,18] where we proposed a
two-level scheduler for Clouds composed of a single datacenter, in this work we extend
the scheduler for operating in federated Clouds. To this end, the scheduler operates at
three levels. Firstly, by means of a policy that operates at the broker level, datacenters
are selected according to their network interconnections and latencies. Indeed, the net-
work latencies among datacenters can contribute to negatively a
ff
ect the response time
delivered to the user. We consider three policies, Lowest-Latency-Time-First (LLTF),
First-Latency-Time-First (FLTF), and Latency-Time-In-Round (LTIR). Then, at the in-
frastructure level, we have explored ACO and PSO for allocating the VMs into the
physical resources of a datacenter. To allocate the VMs into hosts, each scheduler must
make a di
ff
erent number of “queries” to hosts to determine their availability upon each
VM allocation attempt. These queries are actually messages sent to hosts over the net-
work to obtain information regarding their availability. The number of queries to be per-
formed by each algorithm and the latencies of datacenters also influence the response
time to the user. Finally, at the VM level, PSE-jobs are assigned to the preallocated
VMs by using FIFO, as in [18]. Briefly, in this paper we include the broker level and
evaluate how decisions taken both at the broker level and infrastructure level influence
the response time.
Simulated experiments performed with job data extracted from a real-world PSE [6]
involving a viscoplastic problem suggest that the SI schedulers at the infrastructure
level, in combination with these policies at the broker level and FIFO at the VM level,
deliver competitive performance with respect to the response time. Experiments were
performed by using the CloudSim [2] simulator. To set the basis for comparison, and
since VM scheduling is highly challenging and heavily contributes to the overall perfor-
mance in Cloud scheduling [20], we used the same three policies at the broker level and
FIFO at the VM level in combination with a scheduler based on Genetic Algorithms
(GA) [1].
The rest of the paper is as follows. Section 2 gives some background necessary to
understand the concepts underpinning our scheduler. Section 3 surveys relevant related
works. Section 4 presents our proposal. Section 5 presents the experimental evaluation.
Section 6 concludes the paper and discusses future prospective extensions.
ff
