What is the role of the chain-based interaction mechanism in CDynP, and how does it help in generating a near-optimal global parallel strategy without centralized control?
The chain-based interaction mechanism is the core of CDynPs decentralized optimization. It enables the assembly of a near-optimal global strategy by having each data center (DC) in a logical chain perform a local, privacy-preserving optimization and pass only a compressed summary to the next hop.
The process, illustrated in Figure 3, works as follows: The first DC in the chain performs dynamic programming (DynP) on its expanded layer range and compresses the results into a State Summary (SS) containing only minimal stage times, hiding internal GPU allocations.
The next DC receives this SS, uses it to initialize its local DynP process (accounting for the upstream pipeline cost), and then performs its own optimization. This "inherit-plan-compress-forward" cycle repeats along the chain. Finally, the tail node identifies the best global strategy, and a backtracking process propagates the optimal cut-points upstream. This mechanism eliminates the need for a central controller with full system knowledge. Each DCs local decision, guided by the inherited SS, contributes to a globally efficient solution, as proven to be equivalent to a full-information enumeration under the privacy constraints.
