WCET analysis of the shared data cache in integrated CPU-GPU architectures

By taking the advantages of both CPU and GPU as well as the shared DRAM and cache, the integrated CPU-GPU architecture has the potential to boost the performance for a variety of applications, including real-time applications as well. However, before being applied to the hard real-time and safety-critical applications, the time-predictability of the integrated CPU-GPU architecture needs to be studied and improved. In this work, we study the shared data Last Level Cache (LLC) in the integrated CPU-GPU architecture and propose to use an access interval based method to improve the time-predictability of the LLC. The results show that the proposed technique can effectively improve the accuracy of the miss rate estimation in the LLC. We also find that the improved LLC miss rate estimations can be used to further improve the WCET estimations of GPU kernels running on such an architecture.