Information Technology Reference
In-Depth Information
to 50.2% for the makespan and up to 32.0% for the energy consumption. When
comparing the online batch and oine approaches of the best algorithms, results
show the online batch MaxMin is up to 32.3% more accurate than the oine
MaxMin for the makespan, and the online batch MaxMIN is up to 18% more
accurate than the oine MaxMIN for the energy consumption.
Table 4 shows the number of problem instances in which each algorithm is able
to compute the most accurate schedule for each objective. It can be seen that the
previous results hold. The most accurate heuristic is the oine MaxMin when no
error level is considered. The online batch MaxMin is the most accurate for the
makespan objective when higher error rates are considered, and the online batch
MaxMIN is the most accurate for the energy consumption objective also when
higher error rates are considered. Although the online greedy algorithms are able
to compute competitive schedules in average, they are not able to compute the
most accurate result for any problem instance.
7 Conclusions and Future Work
This work presented a formulation for the energy-aware scheduling problem con-
sidering uncertainties in the execution time of the tasks and in the energy con-
sumption of the computing infrastructure. We analysed three real-world task
workloads and proposed a workload generation model considering uncertainties.
We also conducted empirical evaluations to validate and extend our previously
proposed energy consumption model to consider uncertainty values.
In order to analyse the impact of these uncertainty values we evaluated a set of
scheduling algorithms considering different scheduling approaches. Some of these
scheduling approaches being better fitted to cope with uncertainties than others.
Results show the uncertainty values in real-world scenarios significantly affects
the accuracy of the scheduling algorithm, hence considering these uncertainty
values may improve the accuracy of a scheduling algorithm.
In future work, we propose to extend our mathematical model to consider
parallel non-independent tasks and to characterize the energy consumption of
tasks which are not entirely CPU-bound, allowing us to model even more realistic
problem instances and to take advantage of technologies such as DVFS. We
will work on improving the accuracy of our proposed scheduling algorithms and
compare them with some well-known commercial batch scheduler, e.g. Maui.
Acknowledgment.
The work of S. Iturriaga, S. Garcıa and S. Nesmachnow is
partlysupportedbyANII(projectFSE 2013 1 10974)andPEDECIBA,Uruguay.
References
1. Ahmad, I., Ranka, S.: Handbook of Energy-Aware and Green Computing.
Chapman & Hall/CRC (2012)
2. Ali, S., Maciejewski, A., Siegel, H., Kim, J.: Measuring the robustness of a resource
allocation. IEEE Trans. Parallel Distrib. Syst. 51(7), 630-641 (2004)
