Title: Optimization of a Data Race Detector

A data race condition occurs when multiple processes of a parallel program with a shared memory space access the same memory location in an unpredictable order, and at least one write access is involved. This makes the final value stored at the memory location and/or the value read by each process unpredictable. Sometimes a data race is intentional or harmless, but often it is a serious problem caused by a programming error. We have improved an application for data race detection. This application analyzes the data accesses of a multi-threaded or parallel program by inserting, at run time, instrumentation code in front of each instruction of the program that accesses memory, and by intercepting all synchronization operations. This slows down execution considerably. Our task was to optimize this application as much as possible to speed up the execution. We started with a data race detector that could only examine a very short execution span of a multithreaded program before the detector used up all memory and had to give up. Even for such a short execution span this examination took a very long time. We have solved the memory usage problem by combining the registered memory accesses in the data race detector to free up memory, without loss of detection accuracy, and by optimizing almost all algorithms in the data race detecto