| Summary: | Description The paper describes a stateful in-network load balancer for cache objects. In general, a tradeoff arises in in-network caching due to the dynamicity of the requests and the popularity of the objects.The paper describes a time-to-live in-network caching system that is well equipped to accelerate federated learning through in-network model aggregation and caching. Relation to 5G-IANA The described system denoted Lethe is ready for DML integration, i.e., either during training or inference it can be used to store/process ML models in the network instead of the aggregation VNF. It would then expedite training through in-network aggregation as well as inference as it would cache models near to the vehicles /OBUs. Lethe efficiently redirects the request to the relevant ML model, whether it is cached or stored in the database, resulting in reduced response times. Note that caching popular ML models helps reduce memory overhead at the Edge by dividing up model storage between the data network and the edge.
|
|---|