Performance Analysis and Optimization of Parallel Discrete Event Simulation on Intel's Knights Landing Processors

Ali Eker, State University of New York at Binghamton
Intel Optane persistent memory 200 series

Abstract:  We investigate the performance of Parallel Discrete Event Simulation systems on Intel’s KNL processors for thread scheduling and Global Virtual Time algorithms. We demonstrate that the GVT algorithm based on barrier synchronization is a better choice for communication-dominated models, while the asynchronous Mattern’s GVT performs better for computation-dominated scenarios. Based on our findings, we propose a Controlled Asynchronous GVT algorithm that selectively adds synchronization to Mattern-style GVT based on simulation conditions. We also propose a new paradigm for managing PDES threads that we call Demand-Driven PDES. The key idea is to identify the threads that have no events to process and de-schedule them from the CPU until they receive a message requiring event processing. These inactive threads are also excluded from participation in the GVT computation, accelerating that process. We demonstrate that on a KNL processor, DD-PDES significantly outperforms the traditional simulation equipped with the best currently proposed GVT algorithms.

Please use this link to attend the virtual seminar: