Introduction
Fog computing aims to bridge the gap between centralized cloud data centers and end-user devices by placing compute resources closer to the data source. However, this paradigm faces a significant resource management hurdle: function placement. Edge nodes typically possess very limited compute and storage capacity. When a node is overwhelmed by requests, it must decide which serverless functions (FaaS) to run locally to maintain low latency and which to offload to the cloud.
Existing solutions often rely on centralized schedulers, which struggle to scale in highly distributed, dynamic fog environments. Furthermore, there is often no clear economic model to prioritize competing developer requests when resources are scarce. This paper proposes a decentralized, auction-based mechanism to manage these trade-offs autonomously.
Decentralized Auction Mechanism
The authors propose a scheme where each fog node acts as an independent auctioneer. This removes the need for a global "brain" and allows the system to scale naturally. The mechanism relies on a two-part bidding system:
- Storage Bid ($/s): The price a developer is willing to pay to keep function code stored on a specific node.
- Execution Bid ($/execution): The price a developer is willing to pay for a single run of that function.
Nodes make local decisions based on these bids. If disk space is exhausted, the node evicts functions with the lowest storage bids. When processing capacity is reached during request bursts, the node sorts incoming requests by their execution bids, processing the highest bidders locally and pushing the remainder "upstream."
System Architecture
The hierarchy modeled in this research consists of a path from the edge to the cloud, accounting for network latency and node-specific capacities.
As shown in Figure 1, the system is structured to allow multi-tier offloading. Requests originate from client devices and hit the edge nodes. If the edge node cannot satisfy the request (due to storage or compute constraints), it is passed to intermediary nodes in the core network, and finally to the cloud, which serves as a "safety net" with infinite capacity but higher latency.
Experimental Evaluation
To evaluate the mechanism, the authors developed faas4fogsim, a custom simulation tool. They utilized synthetic data with bids randomly sampled from a uniform distribution [50; 150] and varying request loads from 100 to 2000 requests per second (REQ/S).
Key Results: Price Sensitivity and Latency
A primary finding is the system's inherent price sensitivity. As the request load increases beyond the edge node's capacity (approximately 166 REQ/S), the auction successfully prioritizes high-value requests at the edge.
As illustrated in Figure 4, the average price of functions executed at the edge rises significantly as the load increases. This confirms that the auction naturally pushes lower-paying, less latency-sensitive tasks toward the cloud, where prices remain stable. Conversely, developers requiring ultra-low edge latency must bid higher to secure local resources.
Additionally, the study found that as the system becomes overloaded, overall request latency increases linearly. This occurs because lower-paying requests are forced to travel further up the network hierarchy to find available compute power. Similarly, once unique executables exceed edge storage capacity, the node evicts low-bid functions, increasing the latency for those specific functions as they are forced to run in the cloud.
Limitations and Future Work
While the decentralized auction provides a scalable solution, several challenges remain:
- Static Bidding: The current model uses fixed bids; future iterations could involve intelligent agents that adjust bids based on real-time network congestion.
- Lack of Hard Guarantees: The model is "best effort" and lacks hard Service Level Agreements (SLAs), meaning a high bidder could still be displaced if a higher bidder appears.
- Data Placement: The paper focuses on function code placement, but co-managing the placement of large datasets required by those functions remains a significant open challenge.
By treating fog resources as a market, this research provides a foundation for autonomous, revenue-maximizing resource allocation in the next generation of distributed serverless platforms.
Summary updated