Discriminative coherence: Balancing performance and latency bounds in data-sharing multi-core real-time systems

Tasks in modern multi-core real-time systems share data and communicate among each other. Nonetheless, the majority of published research in real-time systems either assumes that tasks do not share data or prohibits data sharing by design. Only recently, some works investigated solutions to address this limitation and enable data sharing; however, we find these works to suffer from severe limitations. In particular, approaches that bypass private caches to avoid coherence interference altogether suffer from significant average-case performance degradation. On the other hand, proposed predictable cache coherence protocols increase the worst-case memory latency (WCL) quadratically due to coherence interference. In this paper, by carefully analyzing the scenarios that lead to high coherence interference, we make the following observation. A protocol that distinguishes between non-modifying (read) and modifying (write) memory accesses is key towards reducing the effects of coherence interference on WCL. Accordingly, we propose DISCO, a discriminative coherence solution that capitalizes on this observation to balance average-case performance and WCL. This is achieved by disallowing modified data in private caches, and hence, the significant coherence delays resulting from them are avoided. In addition, DISCO achieves high average performance by allowing tasks to simultaneously read shared data in the private caches. Moreover, if the system supports the distinction between private and shared data, DISCO further improves average performance by allowing for the caching of private data in cores’ private caches regardless of whether it is modified or not. Our evaluation shows that DISCO achieves 7.2× lower latency bounds compared to the state-of-the-art predictable coherence protocol. DISCO also achieves up to 11.4× (5.3× on average) better performance than private cache bypassing for the SPLASH-3 benchmarks.