Predicting the Performance and the Power Consumption of MPI Applications With SimGrid

Franz C. Heinrich (1, 2) , Alexandra Carpen-Amarie (3) , Augustin Degomme (4) , Sascha Hunold (3) , Arnaud Legrand (1, 2) , Anne-Cécile Orgerie (5) , Martin Quinson (5)

1 LIG - Laboratoire d'Informatique de Grenoble
2 POLARIS - Performance analysis and optimization of LARge Infrastructures and Systems
3 TU Wien - Vienna University of Technology = Technische Universität Wien
4 Unibas - Université de Bâle = University of Basel = Basel Universität
5 MYRIADS - Design and Implementation of Autonomous Distributed Systems

The past decade witnessed a rapid development of powerful but energy-hungry parallel and distributed systems, making energy efficiency of large data centers an important optimization goal. Simulation is a popular approach for studying the behavior of HPC applications in a variety of scenarios. However, simulators are infrequently able to provide faithful performance predictions of applications and typically lack the capability of providing details about the energy consumption of the simulated platforms, especially when comprised of multi-core architectures. Furthermore, studying the impact of different application configurations on energy consumption is a difficult task as only few platforms are equipped with proper power measurement devices. In this paper, we present an extension of the SimGrid simulation toolkit that addresses these challenges. We firstly introduce a model for application energy consumption that supports dynamic voltage/frequency scaling (DVFS) of simulated processors. Secondly, we discuss means to account for coarse-grain memory effects in multi-core architectures. The advantages of our approach, compared to cycle-level simula-tors, are faster simulation run times and enhanced scalability with, provided the target platform is correctly modeled, a retained excellent accuracy. We discuss our model in detail and demonstrate how it can be instantiated by profiling different applications during the calibration phase. Finally, the proposed simulator is validated through an extensive set of experiments with common HPC benchmarks.

Mots clés: SMPI; SimGrid; HPC; Energy Prediction; Reproducible Research

Domaines: Calcul parallèle, distribué et partagé [cs.DC]

Franz C. Heinrich, Alexandra Carpen-Amarie, Augustin Degomme, Sascha Hunold, Arnaud Legrand, et al.. Predicting the Performance and the Power Consumption of MPI Applications With SimGrid. 2017. ⟨hal-01446134⟩