The Journal of Supercomputing
| Hi-index | 14.98 |
The work presents a distributed hardware-level barrier mechanism for n-dimensional mesh-connected MIMD computers, called Distributed Virtual Bit-Slice Synchronizer (DVBSS). The proposed mechanism is structured around an m-bit dedicated control network, whose topology is a directed mesh-embeddable graph, with an additional m-bit-wide wraparound connection. By using a specific virtualization scheme making it possible to have p virtual m-bit barrier networks superposed on a physical one, the DVBSS model allows to synchronize more than m barrier groups. To minimize synchronization latency, the DVBSS scheme uses a distributed circulating wave clocking (DCW-clocking) technique to switch between virtual barrier networks in a pipeline fashion. The DVBSS scheme is shown to be general, configurable, and MPI-compatible. Unlike proposed distributed hardware barriers, and hardware tree-based schemes, the DVBSS mechanism accepts dynamically defined (possibly overlapping) barrier groups of arbitrary size and shape, allowing noncontiguous group member allocations.