Managing RFID events in large‐scale distributed RFID infrastructures

As RFID installations become larger and more geographically distributed, their scalability becomes a concern. Currently, most RFID processing occurs in a central location, gathering tag scans and matching them to event‐condition‐action (ECA) rules. However, as the number of scans and ECA rules grows, the workload quickly outpaces the capacity of a centralized processing server. In this paper, we consider the problem of distributing the RFID processing workload across multiple nodes in the system. We describe the problem, and present an overview of our approach. We then formulate two decision models for distributing the processing across the system. One generates an optimal allocation based on global awareness of the state of the system. This problem is $\mathrm{𝒩𝒫}$ ‐hard and assumes that bandwidth and processing resource availability is known in a central location, which is unrealistic in real scenarios. Thus, we use this model as a theoretical optimal model for comparison purposes. The second model generates a set of local decisions based on locally‐available processing and bandwidth information, which takes much less information into account than the global model, but still produces useful results. We describe our system architecture, and present a set of experimental results that demonstrate that (a) the global model, while providing an optimal allocation of processing responsibilities, model does not scale well, requiring hours to solve problems that the localized model can solve in a few tens of seconds; (b) the localized model generates usable solutions, differing from the optimal solution on average by 2.1% for smaller problem sizes and at most 5.8% in the largest problem size compared; and (c) the localized approach can provide runtime performance near that of the global model, within 3‐5% of the global model, and up to a 55% improvement in runtime performance over a (uniform) random allocation.