Resumen
In scientific domains such as high-energy particle physics and genomics, the quantity of high-speed data traffic generated may far exceed the storage throughput and be unable to be in time stored in the current node. Cooperating and utilizing multiple storage nodes on the forwarding path provides an opportunity for high-speed data storage. This paper proposes the use of flow entries to dynamically split traffic among selected neighbor nodes to sequentially amortize excess traffic. We propose a neighbor selection mechanism based on the Local Name Mapping and Resolution System, in which the node weights are computed by combing the link bandwidth and node storage capability, and determining whether to split traffic by comparing normalized weight values with thresholds. To dynamically offload traffic among multiple targets, the cooperative storage strategy implemented in a programmable data plane is presented using the relative weights and ID suffix matching. Evaluation shows that our proposed schema is more efficient compared with end-to-end transmission and ECMP in terms of bandwidth usage and transfer time, and is beneficial in big science.