A Sensor Distribution Algorithm for FPGAs with Minimal Dynamic Reconfiguration Overhead

Thermal monitoring of a design plays a vital role to ensure safe and reliable thermal operating conditions. Thermal monitoring by employing thermal sensors is a popular technique for assessing thermal behavior and providing directives for run-time preventive measures. Field Programmable Gate Arrays (FPGAs) offer a wide range of functionalities and therefore static placement of thermal sensors during fabrication process is inefficient and inappropriate for such resource constrained devices. In such cases, dynamic reconfiguration is shown to be effective for thermal monitoring because the resources required to instantiate sensors and their exact placement might conflict with the placed and routed circuit. However, instantiating a large number of sensors will incur significant reconfiguration overhead. Our goal is to assist the dynamic reconfiguration based sensor placement approach such that it becomes more efficient. In this paper we present our Longest Edge Rectangular Bisection algorithm, that minimizes the number of required sensors and controls their distribution on the reconfigurable fabric. Our algorithm effectively monitors the design and maintains a given accuracy level in temperature sensing. For a set of hotspots distributed across different logic array sizes, our algorithm reduces the number of required sensors by 69.5% and 17.5% on an average, when compared to a uniform grid-based placement and a more efficient circular intersection based placement respectively.