Parallelism-Aware High-Performance Cache Coherence with Tight Latency Bounds

In Commercial-Off-The-Shelf (COTS) systems-on-chip, processing elements communicate data through a shared memory hierarchy, and a coherent high-performance interconnect, where the de facto standard to handle shared data is through a coherence protocol. Driven by the extraordinary demands from modern real-time embedded system applications to generate, process, and communicate massive amounts of data, recent efforts aim to ensure timing predictability while integrating cache coherence in multi-core real-time systems. However, we observe that most of these efforts compromise system average performance upon offering predictability guarantees. Motivated by this observation, this work proposes an arbiter aimed at providing a predictable, coherent shared cache hierarchy solution, yet with a negligible performance degradation compared to COTS solutions. We achieve this goal by adopting a high-performance-driven architecture including a split-transaction bus and bankized shared cache. In addition, all accesses are arbitrated through a global ordering mechanism. Our proposed arbiter operates alongside conventional coherence protocols without requiring any protocol modifications. Furthermore, we leverage the Duetto reference model by pairing the proposed arbiter and a high-performance arbiter. We evaluate our solution based on both synthetic and SPLASH-3 benchmarks, showing that we can significantly outperform the state-of-the-art in predictable cache coherence, while offering a COTS-level performance