Workload scheduling analysis in geophysical numerical methods

Abstract

The tasks organization among computational nodes affects the performance of the program. In computationally expensive applications such as geophysical problems, the impact is more significant. The imbalance caused by inefficient task scheduling can generate an application non-scalable. Therefore this work investigates the workload scheduling in geophysical methods. Three types of workload scheduling will be compared: centralized dynamic (CD), decentralized static (DS), and decentralized dynamic (DD). The CD was implemented with the technique master-slave, where the master node is responsible for the distribution of the tasks to other nodes called slaves. The DS was implemented with an equal division of tasks before the execution of them. The DD implements the work-stealing method proposed by Assis et al. (2019), where an idle node can steal the tasks of an overloaded node. The principal geophysical method used was 2D Full waveform inversion (FWI) with the acoustic wave. To analyze the performance of workload scheduling methods, we employ a synthetic velocity model and present speedup, efficiency, and load distribution plots generated with different model sizes and different quantity of nodes. The FWI code and the workload scheduling methods were implemented in C with distributed memory parallelization and using the message passing interface (MPI) library. With the results of the workload scheduling methods in FWI, it was applied the workstealing (because this technique presented the more effective performance) in another geophysical problem: Least-squares migration (LSM). We used the LSM with DS implemented by Chauris and Cocher (2017) to compare with the LSM with work-stealing. The LSM code and DS used in this problem were implemented in Fortran and the workstealing in C. The communication between the nodes was implemented using MPI. To analyze the performance of workload scheduling methods in LSM, we used the marmousi velocity model.