Predictable Shared Memory Resources for Multi-Core Real-Time Systems

A major challenge in multi-core real-time systems is the interference problem on the shared hardware components amongst cores. Examples of these shared components include buses, on-chip caches, and off-chip dynamic random access memories (DRAMs). The problem arises because different cores in the system interfere with each other, while competing to access the shared hardware components. It is a challenging problem for real-time systems because operations of one core affect the temporal behaviour of other cores, which complicates the timing analysis of the system. We address this problem by making the following contributions. 1) For shared buses, we propose CArb, a predictable and criticality-aware arbiter, which provides guaranteed and differential service to tasks based on their requirements. In addition, we utilize CArb to mitigate overheads resulting from system switching among different modes. 2) For the cache hierarchy, we address the problem of maintaining cache coherence in multi-core real-time systems by modifying current coherence protocols such that data sharing is viable for real-time systems in a manner amenable for timing analysis. The proposed solution provides performance improvements, does not impose any scheduling restrictions, and does not require any source-code modifications. 3) At the shared DRAM level, we propose PMC, a programmable memory controller that provides latency guarantees for running tasks upon accessing the off-chip DRAM, while assigning differential memory services to tasks based on their bandwidth and latency requirements. In addition to PMC, we conduct a latency-based analysis on DRAM memory controllers (MCs). Our analysis provides both best-case and worst-case bounds on the latency that any request suffers upon accessing the DRAM. The analysis comprehensively covers all possible interactions of successive requests considering all possible DRAM states. Finally, we formally model request interrelations and DRAM command interactions. We use these models to develop an automated validation framework along with benchmark suites to validate and evaluate PMC and any other MC, which we release as an open-source tool.