Dynamic race detection in parallel programs [Elektronische Ressource] / von Ali Jannesari Ladani

Dynamic Race Detectionin Parallel ProgramsZur Erlangung des akademischen Grades einesDoktors der Ingenieurwissenschaftender Fakult¨at fur¨ Informatikdes Karlsruher Instituts fur¨ Technologie (KIT)genehmigteDissertationvonAli Jannesari LadaniDezember 2010Tag der mundlic¨ henPrufung:¨ 03. November 2010Erstgutachter: Prof. Dr. Walter F. TichyZweitgutachter: Prof. Dr. Andreas Zeller(Universit¨at des Saarlandes) KIT-UniversityoftheStateofBaden-WuerttembergandNationalResearchCenteroftheHelmholtzAssociation www.kit.eduIIAbstractRecent hardware developments have pushed parallel computing out of the nicheof numeric applications into the mainstream. Unfortunately, parallel programsmay contain synchronization defects, a class of defect which is difficult to de-tect. A significant source of these defects is the phenomenon of data races, i.e.,unsynchronized accesses to shared data. Since parallel programs are schedule-dependent,reproducingdataracesisoftendifficult. Programsencounteringdataraces often do not crash immediately, resulting in mysterious and unpredictablebehavior.Currently, available tools tend to miss many data races, or to produce an over-whelming number of false alarms, regardless of whether they are based on staticanalysis or dynamic analysis. Both types of analysis also have their own specificproblems.