As part of our ongoing work in intelligent distributed storage, we evaluated the performance of our Deep Reinforcement Learning (DRL) mechanism during the inference stage, comparing it against an optimal solver across various optimization objectives.
🧠Figure 26 illustrates the allocation of storage fragments between edge and cloud infrastructure layers. Key insights include:
âś… The DRL agent closely mirrors the optimal solver across all objectives, showcasing strong decision-making capabilities.
⚡ When optimizing latency (both storage & retrieval), the DRL agent smartly favors edge resources to leverage their low-latency advantage.
💰 For objectives like cost and availability, a strategic shift towards the cloud layer is observed—aligning well with economic and reliability considerations.
📉 Some miss-allocations remain—highlighting areas for future improvement. Notably, during retrieval cost optimization, the optimal solver better utilized the responsive edge layer compared to our DRL agent.
This analysis validates the potential of DRL in real-world distributed systems, while also guiding future enhancements.
#AI #ReinforcementLearning #EdgeComputing #CloudComputing #DistributedSystems #MachineLearning #PerformanceEvaluation #AllegroProject #TechForGood
