Profile directed systematic testing of concurrent programs

Abstract

Runtime data is a rich source of feedback information which can be used to improve program analysis. In this paper, we proposed a Profile directed Event driven Dynamic AnaLysis (PEDAL) to effectively detect concurrency bugs. PEDAL identifies important schedule points with the help of profiling data, and generates a reduced set of schedule points where preemptions could happen. The reduced preemption set is then used to direct the search for erroneous schedules. PEDAL is evaluated on a set of multithreaded benchmark programs, including MySQL, the industrial level database server application. Experimental results show that PEDAL is both efficient and scalable, as compared with several existing analysis techniques.