Information Technology Reference
In-Depth Information
The main trend on the scientific community in energy-aware scheduling is
based on optimizing the energy consumption of the computing elements since
the processor is the main energy consuming element among the hardware com-
ponents. The processor also offers the most flexible options for energy manage-
ment, such as dynamic voltage and frequency scaling (DVFS), dynamic power
management, slack sharing and reclamation, and other techniques [20].
Many scheduling algorithms are based on assuming that the time required to
perform every task is known in advance, and the planning is performed accord-
ing to that input information. However, that assumption does not hold true in
the case of computational infrastructures, where users submit their jobs to be
executed on heterogeneous computing elements. Accurately predicting the exe-
cution time for individual tasks is a very hard problem, mainly because the actual
execution time depends on many factors including the hardware features, com-
munications and delays due to infrastructure and parallel execution, resource
availability, among others. Estimation models using task profiling and bench-
marking have been proposed since the early 1990's [9,10], but they rely on spe-
cific hardware features and computing models that are not fully reasonable for
nowadays clusters and distributed computing infrastructures. Furthermore, cur-
rent models for predicting the energy consumption do include some unrealistic
approximations about the power utilization, especially in the case of complex
multicore servers [16].
This article presents an empirical evaluation of energy-aware schedulers in
heterogeneous computing (HC) scenarios that consider uncertainties in both
the execution time of tasks and the energy consumption for a given computing
infrastructure. We propose three variants of each of the best energy-aware list
scheduling techniques proposed in our previous work [16]. Then, we analyze
their behavior when addressing specific instances of the energy-aware scheduling
problem in multicore HC systems, accounting for realistic errors in the estimation
of the execution time of tasks, and specific deviations in the power consumption
calculation when using a standard energy model for computing systems.
The main contribution of this article consists in proposing novel scheduling
algorithms and reporting their experimental evaluation performed over realis-
tic workloads and scenarios, validated by in-situ measurements using a power
distribution unit. The empirical results demonstrate that error in real-world sce-
narios have a significant impact on the accuracy of the scheduling algorithms.
Different scheduling approaches were evaluated, and the online approach showed
improvements of up to 32% in computing performance and up to 18% in energy
consumption over the oine approach using the same scheduling algorithm.
The paper is organized as follows. Section 2 describes the energy-awareschedul-
ing problem under uncertainty. A review of related work is presented in Section 3.
The heuristics for energy-aware scheduling in high performance computing
systems are introduced in Section 4, just before the description of our model for
uncertainty in Section 5. The experimental analysis of the proposed heuristics is
reported in Section 6, Finally, Section 7 presents the conclusions and formulates
the main lines for future work.
