Barrier synchronization on a loaded SMP using two-phase waiting algorithms

Little work has been done on the performance of barrier synchronization using two-phase blocking, as the common wisdom is that it is useless to spin if the total number of threads in the system exceeds the number of processors. We challenge this and show that it may be beneficial to spin-wait even if the number of threads is up to double the number of processors, especially if the waiting time is at least twice the context switch overhead (rather than being equal to it). We also characterize the alternating synchronization pattern that applications based on barriers tend to fall into which is quite different from the patterns typically assumed in theoretical analyses.