Two algorithms for barrier synchronization

Abstract

We describe two new algorithms for implementing barrier synchronization on a shared-memory multicomputer. Both algorithms are based on a method due to Brooks. We first improve Brooks' algorithm by introducing double buffering. Our dissemination algorithm replaces Brooks' communication pattern with an information dissemination algorithm described by Han and Finkel. Our tournament algorithm Uses a different communication pattern and generally requires fewer total instructions. The resulting algorithms improve Brooks' original barrier by a factor of two when the number of processes is a power of two. When the number of processes is not a power of two, these algorithms improve even more upon Brooks' algorithm because absent processes need not be simulated. These algorithms share with Brooks' barrier the limitation that each of the n processes meeting at the barrier must be assigned identifiers i such that O<~i<n.

Gossamer: A Lightweight Approach to Using Multicore Machines

“…Barrier synchronization points inside a divide/replicate annotation are implemented using a dissemination barrier [32] for scalability. Barrier synchronization is a spin-waiting operation since the number of filaments participating in a barrier is equal to the number of server threads.…”

Insertion Tree Phasers: Efficient and Scalable Barrier Synchronization for Fine-Grained Parallelism

“…It can be adapted to support dynamic participants, but unfortunately does not scale well. The dissemination barrier scales better, but requires barrier participants to be known upfront [9], which prevents its use as a dynamic barrier.…”

“…However, the general concept behind the tournament barrier is adaptable for our purpose, thus we will briefly describe the algorithm of Hensgen et al [9].…”

Performance Evaluation of View-Oriented Parallel Programming on Cluster of Computers

“…We implement the barrier with the binomial-tree model [11] whose complexity is just O(logN) instead of the O(N) of the original linear model. As a result the barrier time on 64 processes in the cluster we mentioned above is 623.8 microseconds, which is 17% faster than the original implementation.…”