This paper introduces a novel approach to processing data-intensive and delay-sensitive requests. It utilizes a heterogeneous fleet of fog nodes where super fog (SF) nodes (high computational resources local to terminal) are always active and handle incoming requests, while other fog (OF) nodes (moderate resources and away from terminals) remain idle and are only engaged when the SF nodes become saturated. In contrast, traditional fog computing typically employs a homogeneous fleet, with all nodes having similar resources and proximity to terminals. The paper proposes a novel load balancing technique wherein all traffic is initially handled by the SF nodes. Once the SF layer becomes saturated, the load is gradually distributed to the OF nodes, mimicking heat diffusion.

The performance of the heterogeneous SF-OF architecture is evaluated against other prominent techniques, such as a priority-aware scheme implemented in a multi-tier architecture with primary and secondary nodes. These architectures are benchmarked across key metrics, including average provisioning delay, admission rate, number of accepted and failed requests, power and energy consumption, total network cost, and network utilization. The SF-OF architecture outperforms multi-tier designs in terms of successful requests, admission rate, blockage rate, and network utilization. For metrics such as provisioning delay, power consumption, energy usage, and network cost, the SF-OF architecture still demonstrates advantages; however, there is a noticeable bimodal behavior in the performance graphs showing distinct trends before and after load migration begins.

In traditional distributed systems, such bimodal designs can be challenging to reason about and validate, as bugs in the secondary mode often go undetected until that mode is triggered under specific load conditions. One way to ensure the reliability of the secondary mode is to route a small, continuous stream of traffic through the OF nodes at all times. This helps to surface latent bugs and ensures the fallback path remains functional. However, this approach may increase power consumption, as the OF nodes are no longer idle and are kept active even during low-load scenarios.