Delay-insensitive synchronization on a message passing architecture with an open collector bus

The performance of some algorithms, running on a message passing computer, is limited by the high latency of global communications. To increase the performance, a simple open collector bus operated by delay insensitive programs running on each processor can be used. We illustrate this by an example: the constraint algorithm SHAKE as used in Constraint Molecular Dynamics simulation. We present a parallelizable SHAKE algorithm and show how it can be implemented on a ring architecture. On a large ring the use of message passing to synchronize SHAKE iterations may take up to 40% of the total time. We show how the communication time can be reduced by adding a very simple open collector bus, operated by a delay insensitive algorithm. In this way the time spent on the synchronization of SHAKE iterations will be negligible. We want to emphasize that this kind of open collector bus can be used with many delay insensitive algorithms. To show this we mention other possible applications