Non-Blocking Technique for Parallel Algorithms with Global Barrier Synchronization

Abstract

Sharing data among asynchronous processes is considered to be a hard systems problem in multithreaded modern shared-memory multicore systems. Throughout the literature, multiple solutions have been proposed, like the so-called barrier synchronization. A Barrier is a synchronization primitive that provides guarantees that any thread will not continue execution from a given point until all threads have reached that point. This primitive is widely used in different parallel programming models, but it can easily become a hot-spot for performance critical applications due to its global nature as one preempted thread will stop execution of all other threads waiting at the barrier. This paper suggests a technique to change the global nature of barrier synchronization into a non-blocking synchronization model with lock-free thread progression guarantees. The main idea is to exploit algorithm-based memory access patterns to implement self-synchronizable threads to protect concurrent reads and writes in a shared data structure without explicit use of a barrier primitive. To the best of our knowledge, this is the first attempt to provide a different synchronization mechanism based on the algorithm intrinsic characteristics rather than an explicit use of a global barrier in shared-memory architectures. Our experimental results show factors of performance improvement against its global barrier-based algorithm counterpart.