Information Technology Reference
In-Depth Information
strategy capable of achieving time and cost savings in the execution of scientific
workflows.
5 Concluding Remarks
This paper presented a novel strategy for cloud infrastructure autoscaling de-
signed for scientific workflows denominated Spots Instances Aware Autoscaling
(SIAA) strategy. The strategy takes advantage of the better cost-performance
relation provided by spot instances in combination with a heuristic scheduling
method for makespan optimization and reduction of the negative effect of out of
bid failures. SIAA permits a highly ecient use of cloud infrastructures reducing
the workflow makespan and the monetary cost of execution.
Four different real-world scientific workflows were selected for evaluating the
performance of SIAA. Results evidenced that assigning a half of the budget
to spot instances permits SIAA to overcome the state-of-the-art autoscaling
methods in a 14.1%-88.0% of makespan. Results also demonstrated that SIAA
conduced to cost reductions of 21.5% to 43.6%. From the experiments is also
evidenced that SIAA is capable of providing good performance levels regardless
of the number of failures occurring. Results highlighted the importance of (
i
)
determining the adequate proportion of the budget assigned to spot instances as
well as (
ii
) having access to bid price prediction methods with good accuracy,
to improve the workflow makespan.
As part of our future work we plan to study checkpointing techniques for
reducing the time and and money loses derived from failures of large duration
tasks. Checkpointing would also permit a more extensive use of spot instances
without compromising the workflow makespan. A second aspect to investigate
in the future is a method for determining the proper value of the
spotsRatio
parameter considering the characteristics of the application and the quality of
the available bid price prediction method. Finally it is interesting to study the
repercussion of data transfer times and cost during the autoscaling process.
These features are crucial for studying new autoscaling techniques designed for
big data applications like, for example, MapReduce workflows.
Acknowledgements.
The first author wants to thank National Scientific and
Technical Research Council (CONICET), for the postdoctoral fellowship granted.
This research is supported by ANPCyT through project No. PICT-2012-2731.
Also, the financial support provided by SeCTyP-UNCuyo through project No.
M004 is gratefully acknowledged. Finally we want to thank the anonymous re-
viewers who helped improving the quality of this paper.
References
1. Agmon Ben-Yehuda, O., Ben-Yehuda, M., Schuster, A., Tsafrir, D.: Deconstructing
amazon EC2 spot instance pricing. ACM T. Econ. Comput. 1(3), 16 (2013)
2014) (Online accessed June 24, 2014)
