Information Technology Reference
In-Depth Information
Current autoscaling techniques present some limitations that deteriorate the
performance of applications. Some techniques [17,11] lack on adaptivity because
they assume a fixed infrastructure and schedule the tasks in a best effort manner.
Other techniques [2] are constructed upon very simple ad-hoc rules that must
be defined by the user. Finally some techniques do not take advantage of the
better cost-performances offered by spot instances [9]. This work proposes a
novel adaptive autoscaling strategy that overcomes the mentioned limitations
exploiting the advantages of using spot instances and dealing with task failures
intelligently to reduce the time and cost of execution for scientific workflows.
This paper is organized as follows. Next section provides a discussion of the
current advances on managing cloud workflows. Section 3 presents a novel au-
toscaling strategy for scientific workflows. Section 4 discusses the experiments
carried out and analyzes the obtained results. Finally, section 5 concludes this
work and provides future research directions.
2 Related Work
The problem of executing workflows in the cloud has been extensively addressed
over the last years. Several techniques and strategies have been proposed to cope
with the objective of achieving fast and cheap executions. We grouped these
approaches in 3 categories: scheduling on fixed-size infrastructures , rule-based
scaling methods, and autoscaling mechanisms.
Scheduling on fixed-size infrastructures : these strategies rely on heuristic and
metaheuristic methods for a best-effort scheduling considering a predefined cloud
infrastructure [17,11]. Their main limitation is that the infrastructure is kept
unchanged during the entire execution. The lack of adaptability to the variable
workload inherent of workflow applications precludes from taking advantage of
time and cost optimization possibilities.
Rule-based scaling : for addressing the problem of load variability on web ap-
plications (e.g. facebook, vimeo, etc), cloud providers suggest using ad-hoc rule-
based methods to adapt the size of the infrastructure [2]; e.g. “if CPU load
overcomes a certain percentage then acquire x new instances”. However, this
type of rules depend on the characteristics of the application running and might
be unsuitable for experimental applications like scientific workflows.
Autoscaling : these are techniques specially designed for workflow applications
to cope with the problems of scaling and scheduling simultaneously [9]. Under
this category, two strategies denominated Scheduling First and Scaling First
have been proposed. As their names indicate the strategies differ in which of
the phases (scaling or scheduling) are accomplished first. On both cases, the
strategies operate continuously while the applications are running. The men-
tioned strategies use on-demand instances disregarding the use of spot instances
missing important time and cost saving opportunities . As both strategies
present a considerable complexity they are not further discussed due to space
limitations. To obtain a better understanding of the strategies, please refer to
the existing literature [9].
 
Search WWH ::




Custom Search