SWOT Framework Enables Reconfiguration-Communication Overlap in Optical Networks for DML
A recent study introduces SWOT, an optical network framework that is sensitive to demand and employs intra-collective reconfiguration to synchronize network resources with the communication traffic patterns found in distributed machine learning. To mask the latency associated with reconfiguration, SWOT integrates it with data transmission using three primary methods: Heterogeneous Message Splitting, Asynchronous Overlapping, and additional techniques. This research can be accessed on arXiv with the identifier 2510.19322.
Key facts
- Collective communication is critical for scaling distributed machine learning.
- Predictable traffic patterns of DML enable application of optical network technologies.
- Optical networks with reconfigurable topologies promise high bandwidth and low latency.
- Static topologies are inefficient for dynamic communication patterns within CC algorithms.
- Frequent topology reconfiguration incurs significant overhead.
- SWOT is a demand-aware optical network framework proposed in the paper.
- SWOT employs intra-collective reconfiguration to dynamically align network resources.
- SWOT hides reconfiguration latency by overlapping it with data transmission.
- Three key techniques: Heterogeneous Message Splitting, Asynchronous Overlapping.
- The paper is published on arXiv with ID 2510.19322.
Entities
Institutions
- arXiv