Interference resilient PDES on multi-core systems: towards proportional slowdown

摘要: Parallel Discrete Event Simulation (PDES) harnesses the power of parallel processing to improve performance and capacity simulation, supporting bigger models, in more details for scenarios. PDES engines are typically designed evaluated assuming a homogeneous computing system that is dedicated simulation application. In this paper, we first show presence interference from other users, even single process an arbitrarily large environment, can lead dramatic slowdown simulation. We define new metric, which call proportional slowdown, represents idealized target graceful interference. identify some reasons why simulators fall far short slowdown. Based on these observations, design alternative scheduling mapping algorithms better able tolerate More precisely, most resilient will allow dynamic event execution resources (a work pool model). However, model has significant overhead substantially impact locality. Thus, propose locality-aware adaptive dynamic-mapping (LADM) algorithm multi-core systems. LADM reduces number active threads interference, avoiding having disabled due context switching. reduce while maintaining memory locality reducing gap with similar techniques also help situations where there load imbalance or processor heterogeneity.