Eşzamanlı program hatalarının tespiti ve hareketsel bellek yöntemleri ile düzeltilmesi

We propose a technique to improve detection of concurrency errors of multi-threaded C/C++ applications and recovery of these applications from the errors using transactional memory (TM) technology. Transactional memory is an emerging parallel programming model which simplifies parallel program design and implementation, improves performance and protects applications from most concurrency bugs. Our approach uses TM to improve performance of detection of concurrency errors and provides a framework by which legacy C/C++ applications can benefit from concurrency error-freedom. The current concurrent error detection approaches are either too slow or need extra modification to current processor architecture. The slowdown of these techniques stems from a number of reasons: instrumentation used to add them into application, necessary computations needed to detect errors, and cost of proper protection of error-detection-related data used by these techniques. We propose a way to divide the instruction stream of each thread in a multi-threaded application into small transactions. We then use conflict detection to get fine-grain protection of concurrency error detection data to improve performance. We use transaction write buffers and rollback mechanism to recover from concurrency errors, data races in particular, and impose extra protection on erroneous data or portions of concurrent program. Our approach works well on a number of multi-core benchmark applications and shows a significant performance improvement of concurrent error detection over conventional means. These improvements are encouraging initial results for the industrial usage of the proposed approach.