Abstract
The technology described in this paper relates to the use of disjoint Hamiltonian cycles to support asymmetric communication patterns effectively while maintaining desirable latency control properties of direct-connect topologies. The technique used in this paper dynamically re-maps fixed disjoint dimensions of a direct-connect topology into a flexible set of disjoint Hamiltonian cycles that can trade network dimensions and the bandwidth-per-dimension to support asymmetric communication patterns. The mapping of the asymmetric network onto Hamiltonian cycles enables dividing the workload across Hamiltonian cycles of equal length.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
N/A, "EFFICIENT MACHINE LEARNING AND SERVING USING DISJOINT HAMILTONIAN CYCLES", Technical Disclosure Commons, (February 03, 2026)
https://www.tdcommons.org/dpubs_series/9272