Lightning in the Dark: Scheduling Non-Clairvoyant Coflows at Wireless Edge

Deploying distributed computing applications to the network edge offers promising solutions to the growing interest in latency reduction and privacy protection. Nevertheless, these systems are often bottlenecked by intricate communication patterns, where collections of parallel flows belonging to the same job, abstracted as coflow, are exchanged across edge devices. To overcome this bottleneck, we introduce LightningFlow, an efficient scheduling system tailored for non-clairvoyant coflows in interference-constrained wireless networks. LightningFlow differentiates inter- and intra-coflow priorities across edge devices through an interference-aware flow preemption mechanism, which simultaneously improves flow parallelism and mitigates starvation, thereby minimizing application-level latency. We demonstrate via extensive experiments that LightningFlow reduces the mean flow time by 1.6−7.0× and the P95 tail latency by 1.7 − 10.6× compared to the state-of-the-art methods.