Information Technology Reference
In-Depth Information
3
Related Work
In this paper we address the scheduling of scientific application in federated Clouds in
order to minimize the response time considering the influence of the network latencies
among datacenters. Our approach di
ers from those presented in literature for federated
Cloud, where the authors have not considered SI-based strategies at the infrastructure
level. In previous works of our own [16,18] we have presented SI-based schedulers
focused on the infrastructure level. However, it is important to note that in these works
the schedulers operate at two levels for Clouds composed of a single datacenter. The
remaining works found in literature are focused on one level and do not evaluate the
three levels such as we propose in this work.
Among these works we can mention [5,10,20]. In [5] the authors summarize some
VM allocation policies based on linear programming for di
ff
erent Cloud federation ar-
chitectures. Then, in [10] scheduling strategies at the broker level based on di
ff
erent
optimization criteria (e.g., monetary cost optimization or performance optimization)
and di
ff
erent user constraints (e.g., budget, performance, VMs types) were proposed.
Moreover, in [20], the scheduler restricts the deployment of VMs according to some
placement constraints (e.g., Clouds to deploy the VMs) defined by the user.
Two works that deserve special attention are [1,4]. In [1], the authors used at the
broker level, a Dijkstra algorithm to select the datacenter with lower monetary cost,
and a GA for allocating the VMs at the infrastructure level. Although in this work
the authors target the broker and the infrastructure levels, the goal was to reduce the
monetary costs without considering the response time. For scientific applications in
general, the response time is very important [18]. Moreover, in [4] the authors proposed
an ACO scheduler based on load balancing to perform e
ff
cient distribution of jobs by
finding the best VM to execute jobs. The aim of this work was minimizing the makespan
and improve load balancing in the VMs. Makespan is the maximum execution time of
a set of jobs. To the best of our knowledge, this is the only work in literature in which
the authors have considered the use of SI for federated Clouds. However, it is important
to note that ACO was implemented at the VM level and not at the infrastructure level.
With respect to works which address the scheduling problem at the infrastructure
level -intra-datacenter- using SI-based strategies as we propose in this work, few e
orts
have been found [17]. However, in these related works, it is important to note that SI
techniques are used to solve the job scheduling problem, i.e., determining how the jobs
are assigned to pre-allocated VMs, and few e
ff
orts have aimed to solve VM scheduling
problems to date [17]. It is worth noting that, from the related works found, most of them
have been proposed for Clouds taking into account only one of the scheduling levels
without considering SI for allocating VMs, or Clouds composed by a single datacenter
where only scheduling of jobs (and not VMs) is addressed. The next Section explains
our approach, which considers the three levels described in Subsection 2.1.
ff
4
Proposed Scheduler
The goal of our scheduler is to minimize the response time of a set of PSE jobs. Re-
sponse time is the period of time between a user makes a request to the Cloud and gets
