The Non-blocking Programming Paradigm in Large Scale Scientific Computations

Non-blocking implementation of shared data objects is a new alternative approach to the problem of designing scalable shared data objects for multiprocessor systems. Non-blocking implementations allow multiple tasks to access a shared object at the same time, but without enforcing mutual exclusion to accomplish this. Since, in non-blocking implementations of shared data objects, one process is not allowed to block another process, non-blocking shared data objects have the following significant advantages over lock-based ones: 1) they avoid lock convoys and contention points (locks). 2) they provide high fault tolerance (processor failures will never corrupt shared data objects) and eliminates deadlock scenarios, where two or more tasks are waiting for locks held by the other. 3) they do not give priority inversion scenarios. As shown in [1, 2], non-blocking synchronisation has better performance in certain application than blocking synchronisation. In this paper, we try to provide an in depth understanding of the performance benefits of integrating non-blocking synchronisation in scientific computing applications.