An agent-based model for analyzing the HPC input/output system

Abstract

High Performance Computing (HPC) applications canspend a significant portion of their execution time doing Input/Output (I/O) operations into files. Improving I/O performancebecomes more important for the HPC community, as parallelapplications produce more data and use more computing resources. One of the methods used to evaluate and understandthe I/O performance behavior of such applications in new I/Osystems or for different configurations is using modeling andsimulation techniques. In this paper, we present a simulationmodel of the HPC I/O system by using Agent-Based Modelingand Simulation (ABMS) based on the functionality of the I/OSoftware Stack. Our proposal is modeled using the concept ofwhite box so that the specific behavior of each of the modulesor layers in the system can be observed. The interaction betweenthe layers of the I/O software stack are analyzed by monitoringthe internal functions using proprietary parallel file system tools.This allows obtaining the functional and temporal characteristicscorresponding to the I/O operations. These characteristics allowedthe design and implementation of a representative model ofI/O system components. Furthermore, measurements are usedto obtain the necessary data sets in the verification, fine-tuningand validation stages. The resulting implementation has shownsimilar behaviors for measured and simulated values when usingthe IOR benchmark with various file sizes.