Performance Monitoring and Analysis Components in Adaptive PDE-Based Simulations 1
Boyana Norris, Ivana Veljković · 2005
As scientists incorporate more sophisticated models into their simulations, software complexity,aswellasthe underlyingcomputational costof thesemodels,are growing rapidly. Performance evaluation and tuning of applications that are largescale both in terms of source code and runtime requirements can be challenging andtime-consumingforscientists.Wehavedevelopedasoftwareinfrastructurefor performance monitoring, performance data management, and adaptive algorithm developmentforcomponentPDE-basedsimulations.Newton-Krylovnonlinearand linear solver components are instrumented for performance monitoring using the TAU performance tools. To reduce the performance monitoring and component adaptation overhead, we employ two databases that serve signiflcantly difierent purposes. The flrst one is created and destroyed during runtime, and stores performance data for code segments of interest, as well as various application-speciflc performance events in the currently running application instance. The second one ispersistentandcontainsperformancedatafromvariousapplicationsanddifierent instances of the same application. It can also contain performance information derivedthrougho†ineanalysisofrawdata.Wedescribeaprototypeimplementation ofthisinfrastructureandshowhowadaptivelinearsolveralgorithmsareemployed in a driven cavity ∞ow simulation code.