Performance Evaluation of the Barrier Synchronization Solutions on the AP1000

The aim of our research on AP1000 is to measure the overhead of some barrier algorithms and analyze how to take advantage of the AP100 architectural properties in order to obtain the best performance. We have measured the performance of various barrier algorithms on the AP1000. Our results were obtained by embedding barrier episodes inside a loop and averaging over a large number of operations. We have implemented five algorithms, two of which are centralized. Another one is a tree algorithm called U-mesh[1] that exploit the properties of the switching technologies found in machines like the AP1000. We also implemented the Dissemination barrier[2] and an algorithm with some characteristics of the U-mesh and Dissemination. The best performance was found in the last two algorithms. The barrier time was three times smaller than the barrier time of the MPI_Barrier collective routine implemented on AP1000[3].