UCC now supports broadcast collectives over strided active sets larger than two ranks within the TL/UCP transport. This feature allows applications and runtimes to define a specific subset of participating ranks - using the active_set.start, active_set.stride, and active_set.size fields - which UCC then maps into a subset rank space to execute the broadcast only across those selected participants. This support is available in the TL/UCP broadcast path, including the knomial and double-binary-tree broadcast algorithms, and removes the previous limitation where active-set broadcast was restricted to a maximum of two ranks.
Last updated: