SPEED: Scalable and Predictable EnhancEments for Data Handling in Autonomous Systems

Scalable and predictable disk I/O management is critical for autonomous applications that must handle realtime sensor data streams. Existing ROS-based architectures for end-to-end autonomous compute pipelines suffer from lack of scalability and performance predictability that can compromise safety. We present SPEED, an approach leveraging a multi-queue architecture and a context-aware I/O scheduler to achieve scalability and performance predictability. We demonstrate SPEED's ability to reduce single I/O latency in rosbag by 84%, and improve by over 1.5 × the scalability of existing camera sensors. Furthermore, SPEED's context-aware I/O scheduler improves the predictability of multitask application pipelines by over a 3.08× reduction in performance standard deviation. Our results demonstrate significant improvements in I/O management, making SPEED well suited for the stringent I/O demands of modern end-to-end autonomous applications that must execute sensor-to-actuator compute pipelines while meeting realtime performance requirements.