Software Transactional Memory, OpenMP and Pthread implementations of the Conjugate Gradients Method - a Preliminary Evaluation

This paper shows the runtime and cache-efficiency of parallel implementations of the Conjugate Gradients Method based on the three paradigms Software Transactional Memory (STM), OpenMP and Pthreads. While the two last named concepts are used to manage parallelization as well as synchronization, STM was designed to handle only the latter. In our work we disclose that an improved cache efficiency does not necessarily lead to a better execution time because the execution time is dominated by the thread wait time at the barriers.