Improving the Execution Time of Industrial Applications through Planned Cache Eviction Policy Selection

Modern industrial applications are demanding high computational power due to the evolution of features and components, such as real-time communication and control, image processing techniques, and security. In the microprocessor space, the high demand for computational power can only be satisfied by the integration of high-performance hardware components. Amongst them, caches and their replacement policies play a major role in preventing costly off-chip memory accesses.In this article, we argue that enabling caches with several eviction policies that can be selected at the software-layer and changed during an application execution can help substantially decrease its execution time (by increasing the cache hit ratio). To demonstrate this, we (1) present the implementation of an open-source cache simulation framework, (2) propose four distinct approaches to identify when to change the policy during the execution, and (3) assess the cache hit rate improvements brought. Experiments show that workloads running with the proposed approaches can feature up to 30% cache miss rate improvement in comparison to using the LRU replacement policy.