CPSLab Cyber-Physical Systems Lab, Boston University

Publications Memory & Shared-Resource Management

Under Pressure: Adaptive Isolation for Visual SLAM Perception Workloads on Contended ARM Hardware

Alessandra Maria Lanz, Matias Ou, Mattia Nicolella, Renato Mancuso

Industrial Session of the 38th Euromicro Conference on Real-Time Systems (ECRTS 2026), July 2026, Lund, Sweden

Read the PDF Publisher page (DOI) Code and artifacts Cite

Figure 1 Controller architecture. During a preliminary solo run, the calibrator records per-phase instruction-retirement rates and writes them to calib.txt. At runtime, the dynamic controller installs uprobes at five pipeline checkpoints, reads sample records from the perf ring buffer, compares observed progress to the calibrated baseline, and issues SIGSTOP/SIGCONT to the interferer.

Abstract

We present a user-space feedback controller for protecting the tracking accuracy of OV 2 SLAM, the visual SLAM component of the Arm AR HUD Industrial Challenge, when it runs alongside competing workloads. The controller monitors the SLAM pipeline at five function-entry checkpoints using Linux uprobes, measures per-frame progress against a pre-calibrated solo-run baseline, and pauses the interfering workload via SIGSTOP/SIGCONT when progress falls below a threshold. We implement and evaluate the controller on a Xilinx ZCU102 development board (quad-core ARM Cortex-A53, 1 MB shared L2 cache) running Ubuntu 22.04 and ROS2 Humble, using a synthetic memory-bandwidth aggressor as the interfering workload. Under uncontrolled interference, mean trajectory RMSE rises to 1.03 m with 3 out of 5 trials diverging. With the controller active, mean RMSE drops to 0.095 m with no trials diverging. All tools are written in C and run entirely in user space, requiring no kernel modifications and no recompilation of the SLAM binary.

Code and hardware artifacts

Cite this paper

@inproceedings{UPROG_ecrts-ic-26,
  title     = {{Under Pressure: Adaptive Isolation for Visual SLAM Perception Workloads on Contended ARM Hardware}},
  author    = {Lanz, Alessandra Maria and Ou, Matias and Nicolella, Mattia and Mancuso, Renato},
  booktitle = {Industrial Session of the 38th Euromicro Conference on Real-Time Systems (ECRTS 2026)},
  year      = 2026,
  month     = jul,
  address   = {Lund, Sweden},
  doi       = {10.4230/LIPIcs.CVIT.2016.23},
  url       = {https://doi.org/10.4230/LIPIcs.CVIT.2016.23}
}

Download .bib