Dynamic Reconfiguration of Two-Level Caches in Soft Real-Time Embedded Systems

Cache reconfiguration is a promising optimization technique for reducing memory hierarchy energy consumption with little or no impact on overall system performance. While cache reconfiguration is successful in desktop-based systems, it is not directly applicable in real-time systems due to timing constraints. Existing scheduling-aware cache reconfiguration techniques consider only one-level cache. It is a major challenge to dynamically tune multi-level caches since the exploration space is prohibitively large. This paper efficiently integrates cache reconfiguration in soft real-time systems with a unified two-level cache hierarchy. We utilize a set of exploration heuristics during our static analysis which effectively decreases the exploration time while keeps the generated profile results beneficial to be leveraged during runtime. Our experimental results have demonstrated 32 - 49% energy savings with minor impact on performance.