Introduction
In the era of autonomous systems, high-performance embedded platforms must manage a mix of safety-critical and non-critical applications. These applications often share hardware resources like memory controllers and L3 caches. On multicore processors, this sharing leads to "multicore timing interference," where non-critical background tasks (e.g., logging) compete for bandwidth with critical tasks (e.g., emergency braking). Uncontrolled interference can cause critical tasks to miss deadlines, compromising system safety.
While hardware-specific bandwidth regulators exist, they often require specialized kernels or hardware modifications, creating barriers for developers using the modular ROS2 (Robot Operating System 2) framework. ROSGuard addresses this by providing a portable, middleware-level regulation mechanism that operates on standard Linux stacks.
System Architecture
ROSGuard implements bandwidth regulation entirely within the ROS2 ecosystem, avoiding the need for kernel-level patches. The architecture consists of three primary components: the Real-Time (RT) Node, the non-Real-Time (nRT) Node, and a centralized Controller Node.
The Controller Node acts as a supervisor, monitoring memory bandwidth usage via the CPU's Performance Monitoring Unit (PMU). It communicates with regulated cores to ensure that nRT nodes do not exceed their pre-allocated bandwidth budgets, thereby protecting the execution of critical tasks on other cores.
The Throttling Mechanism
The core innovation of ROSGuard is its "middleware-level enforcement." Instead of relying on low-level hardware interrupts, it utilizes standard pthread priorities and ROS2 Executors to manage task execution.
When the Controller detects a bandwidth violation, it triggers a high-priority "Throttle" (THR) callback on the offending nRT node's core. Because the THR callback is assigned a higher priority than the node's nominal behavior (NCT), it preempts the application logic. The THR callback then executes an idle loop, effectively "pausing" the noisy neighbor for the remainder of the regulation window. This approach allows for millisecond-level granularity, which the authors demonstrate is sufficient for most industrial ROS2 applications.
Experimental Evaluation
The researchers evaluated ROSGuard on an NVIDIA AGX Orin platform using synthetic benchmarks (Isolbench) and real-world computer vision tasks from the San Diego Vision Benchmark Suite (SD-VBS).
Key Results
- Minimal Overhead: Monitoring bandwidth usage at a 1kHz sampling rate introduced only a ~1.06% performance slowdown, making it highly efficient for real-time use.
- Rapid Response: The system achieved an activation delay—the time to stop a violating node—of less than 200µs for self-sampling and under 100µs for external sampling.
- Effective Protection: ROSGuard provided protection comparable to specialized state-of-the-art solutions like MemGuard, while maintaining significantly higher portability and ease of deployment on standard hardware.
Conclusion
ROSGuard demonstrates that effective bandwidth regulation can be achieved at the middleware level without sacrificing portability. By mapping hardware regulation concepts to ROS2 abstractions, it empowers developers to build safer autonomous systems on standard hardware. Future work will focus on expanding this mechanism to handle contention in heterogeneous resources, such as shared paths between CPUs and GPUs.
Summary updated