On-the-fly detection of data races in OpenMP programs

Efficient Filtering Technique for Reducing Time Overhead of Dynamic Data Race Detection in Multithread Programs

“…The filtering techniques exclude unnecessary monitoring of memory operations, such as read only operations and local variables, to reduce the dramatic overheads of the dynamic analysis and to insert minimal monitoring cods. In our prior work [16], we presented a filtering technique, called Redundant Event Filtering (REF), which ignores repeated accesses to a shared memory location. Considering the first access on a thread segment is well-known that it is important to debug a parallel program, because data races involving the first accesses of each thread segment may affect later accesses and may lead to other newly appeared data races.…”

“…The data race detection technique of [16] can be firmly established by REF. The dynamic data race detection with REF considers only the first access of each access event type (read and write) in a thread segment, if the accesses to a shared memory location are performed with redundant locks.…”

Effective Monitoring Memory Operations for Dynamic Race Detection through Hierarchical Filtering Method

“…On-the-fly methods [7][8][9][10][11][12][13][14][15][16] dynamically detect data races with still less overhead in storage space than other dynamic techniques, because unnecessary information are removed as the detection advances. On-the-fly methods are based on three different analysis methods: lockset analysis, happens-before analysis, and hybrid analysis.…”

“…Lockset analysis [14,16] reports data races of monitored program by checking violations of a locking discipline, and happens-before analysis [7][8][9][10] reports data races between current access and maintained previous accesses by comparing their happens-before relation based on the usage of a logical time stamp, such as vector clocks. The lockset analysis is simple and can be implemented with low overhead.…”

“…It is important to detect data races for debugging multithreaded programs, because they may lead to non-intended results. Detection techniques can be classified into static analysis [3][4][5] which analyzes source codes without any execution and dynamic analysis [6][7][8][9][10][11][12][13][14][15][16] which locates data races from an execution of multithreaded program.…”