A Study of the Performance of Scientific Applications with Dynamic Loop Scheduling under Perturbations

Scientific applications, such as N-body, Monte Carlo, and computational fluid dynamics consist of large loops. These loops contain computationally-intensive operations, resulting in heavy loop bodies. Loop scheduling techniques are used to parallelize such applications. Dynamic loop scheduling (DLS) techniques are used to mitigate variations in loop iterations execution times caused by problem, algorithmic, or systemic characteristics and, therefore, achieve a balanced load execution of scientific applications on high performance computing systems. Such variations are referred to as perturbations and include, decreased delivered computational speed, reduced available network bandwidth, or larger network latencies. The perturbations can also be caused by other applications or processes that share the same resources, or a temporary system fault or malfunction. In this poster, the performance of a computer vision application scheduled using DLS is studied under nine different perturbation scenarios. The application execution is simulated and its performance is analyzed. The evaluation of the simulation results suggests that no single scheduling technique achieves the best overall performance in all the considered scenarios. This work reveals the need for a mechanism to select the best performing scheduling technique based on the system state during execution to achieve improved application performance.