Where should a serverless function run across cloud, fog and edge? A reverse Vickrey auction (GIRAFF), greedy heuristics and a Dueling Double DQN agent compared in the FaaS-Sim simulator. Team project with Richard Gozan and Lynn Siewe.
Key metric
100%
SLAs accepted by the DQN agent, vs 92.8% for GIRAFF · 27-node topology
Mixed workload · 27-node topology
SLAs accepted%avg cost
RL (DQN)100.00.99
Edge-Furthest93.21.08
GIRAFF92.81.22
Edge-First87.61.67
Edge-Ward84.61.48
Average cost per accepted SLA, in thousandths of a euro · project report, table 1
Architecture
Smart Bidding
Components, data flows and trust boundaries of the system.
Scroll the diagram →
Technologies
Python
FaaS-Sim
Vickrey auction
Dueling Double DQN
MDP · action masking
Energy cost model
LaTeX
Design decisions
01
Second-price reverse auction: each node bids its true marginal energy cost; the cheapest wins and is paid the second bid, which makes truthful bidding optimal without batching requests.
02
Cost is the extra energy the function induces, P(u + uₛ) − P(u): a loaded node prices itself out, which yields emergent load balancing.
03
The RL agent only chooses among feasible nodes (action masking), and its reward mirrors the global acceptance − normalized-cost objective; λ₂ = 20 gives the best trade-off.
04
Since every node is an action, learning runs on 27 nodes; the ~700 nodes of Phase I call for multi-agent approaches (MARL).
The problem
In the cloud-to-edge continuum, nodes are heterogeneous and owned by different actors. Placing each function online under latency, CPU and memory constraints, while trading acceptance rate against energy cost, is NP-hard (reduction from two-dimensional bin packing).
What I built
Problem formulation and a marginal energy-cost model P(u) = au² + bu + c. Integration into FaaS-Sim of GIRAFF (second-price reverse auction with breadth-first bid propagation), three heuristics (Edge-First, Edge-Furthest, Edge-Ward) and a Dueling Double DQN agent modeled as an MDP with infeasible-action masking.
Result and trade-off
On ~700 nodes, GIRAFF stays close to the best acceptance rate while costing less than Edge-Furthest and Edge-First. On the reduced 27-node topology, the DQN agent (λ₂ = 20) accepts 100% of SLAs and cuts total cost by 12% versus GIRAFF, with a more balanced spread across edge, fog and cloud.
Scope and limitations
Simulation results, with centralized decisions and full observability. The RL agent is evaluated on 27 nodes because each node is an action: scaling up will require multi-agent approaches (MARL). Energy, fairness across nodes and SLA-violation penalties remain to be evaluated.
