Broadcasting on Meshes with Wormhole Routing

On the Performance of Broadcast Algorithms in Interconnection Networks

“…Most current practical parallel machines, including meshes, employ Dimension-ordered routing, where messages visit network dimensions in a pre-defined order, [2,5,20]. This is due to the fact that this form of routing is simple and deadlock free, resulting in fast and compact routers when the algorithm is implemented in hardware, [5].…”

“…This is due to the fact that this form of routing is simple and deadlock free, resulting in fast and compact routers when the algorithm is implemented in hardware, [5]. To exploit these features, many researchers have concentrated on trying to enhance broadcast algorithms by observing that the communication latency is primarily generated by the number of message-passing steps between the involved processors [2,11,20,21,22]. For instance, Barnett et.…”

“…al. [2] have originally proposed the Recursive Doubling (RD) algorithm. This algorithm requires N 2 log message-passing steps for broadcasting in an N -node mesh.…”

“…This process is implemented recursively until the completion of the broadcast operation. In the absence of message contention, the RD algorithm can take advantage of the pipelining feature of wormhole switching [2,20]. Although the RD algorithm can be implemented on a multi-port router architecture, it is often unable to take advantage of this architecture model, resulting in more message-passing steps being than necessary.…”

“…Upon receiving the message, each selected side sends the message to the opposite side in its partitioning set, covering the rest of the nodes of the system in parallel [28]. While most previous broadcast algorithms, including the DB algorithm described above, have been discussed in the context of deterministic routing [2,11,20,21,22], this section focuses on the Adaptive Broadcast (AB for short) algorithm presented in [27].…”