AI Architecture & Systems

Collective–Topology Mapping and Communication Offload

SourcesSurvey §4.4, especially its topology-mapping and in-network-reduction discussion.

4.23.1Concurrent Ring, Tree, and Pairwise Exchange on Direct/Switched Fabrics#

4.23.2Dimensional Decomposition and Multi-Ring Embedding on Mesh/Torus#

4.23.3Intra-Group–Inter-Group–Intra-Group Scheduling on Dragonfly/Boardfly#

4.23.4Hierarchical Locality and Link Sharing in UB-Mesh/Maia#

4.23.5NVLS/CollNet-Style Reduction Offload#

4.23.6Data Movement in AllGather/All-to-All and the Limits of Reduction Offload#

4.23.7Communication Placement, Link Contention, and Compute–Communication Overlap#