Coordinating CCL Re-Optimization and WAN Reconfiguration for Cross-Regional AI Acceleration

Many emerging Artificial Intelligence (AI) workloads rely on cross-regional distributed training, where Collective Communication Libraries (CCLs), such as AllGather, AllReduce, and AlltoAll, operate across multiple datacenters (DCs) interconnected by wide-area networks (WANs). However, while WAN routing can adapt to traffic dynamics at sub-second to second timescales, CCL logical-topology reoptimization is typically applied far more slowly, often every tens or hundreds of collectives. Due to this intrinsic timescale mismatch, CCL performance (e.g., CCL 'completion time') is affected by bandwidth variability in the WAN, motivating us to investigate the individual influence on performance of WAN reconfiguration and CCL reoptimization. In this work, we study the effect of applying fast WAN path reconfiguration and slow CCL logical-topology reconfiguration in cross-regional AI training. Rather than proposing a new coordination policy, we provide an initial evaluation of how these existing mechanisms behave when used individually or jointly under dynamic WAN conditions. By analyzing their combined operation, our study also offers practical insights for rapidly deployable cross-regional training systems, since it builds directly on mechanisms already supported by modern CCLs and WAN controllers. Using a simulated inter-DC environment, we observe that combined WAN reconfiguration and CCL re-optimization can significantly reduce AllGather completion time compared to applying only WAN or only CCL adjustments. These results highlight the tight coupling between CCL behavior and WAN dynamics and motivate the development of more principled coordination mechanisms for practical deployment of cross-regional CCLs.