Background

In recent years, automotive and aerospace systems have started to use high-performance System-on-Chip platforms and operating systems. These environments often run multiple applications with different levels of criticality at the same time. Such systems are called mixed-criticality systems.

When multiple applications are integrated on one platform, real-time performance and safety become difficult to guarantee. Applications may interfere with one another through shared CPUs, memory, GPUs, storage devices, and network bandwidth.

The challenge

Contention on shared resources can delay important tasks. In severe cases, deadlines are missed, faults propagate across applications, or new attack paths are opened.

In systems that control vehicles or other critical equipment, such delays therefore affect safety as well as performance.

Left: high-priority, low-priority, and interrupt processing access the shared resources (SoC, GPU, storage) through the communication middleware and the OS, and their accesses interleave. Right: a schematic distribution of access time, where most accesses cluster around the mean but a few take much longer.
Processes competing for shared resources (left) and a schematic distribution of access time (right). Mean and variance matter for efficiency, while for real-time behavior it matters that the rare long delays can be predicted.

What we do

We measure and analyze how real-time applications use shared resources, and identify the causes of delay. The scope covers CPUs, memory systems, GPUs, storage, and communication paths.

We also work across the stack to improve real-time behavior: software-level isolation, changes to operating systems and drivers, and hardware support for tighter timing control.

Achievements so far

I have worked on this area since my doctoral research. Real-time scheduling algorithms for temporal protection (2007) and a flexible scheduling framework for integrating real-time applications (Computer Software, 2008) laid the groundwork for letting tasks of different criticality coexist safely, and the results were released as the open-source scheduling simulator schesim (2012).

The work then expanded to multi-core and many-core platforms: a comparative analysis of RTOS and Linux scalability on an embedded many-core processor (Journal of Information Processing, 2018), the generic profiling infrastructure esprof (Computer Software, 2020), a performance evaluation of mixed-criticality systems on PREEMPT_RT Linux (Journal of Information Processing, 2023), and energy-efficient DVFS scheduling (IEEE Access, 2019).

We also develop educational platforms: EV3RT for LEGO Mindstorms (2017), which received a software paper award, and its successor SPIKE-RT (2022) are widely used in embedded-systems education in Japan. Recent work explores next-generation platforms, including mixed-criticality service orchestration for software-defined vehicles using Lingua Franca (IEEE Embedded Systems Letters, 2025), a Linux-compatible unikernel for constrained devices (OSPERT 2025), and GLACI, a code instrumentation tool for OpenGL (ICONS 2025, Best Paper Award).

Why this matters

When this works, the dozens of computers in a car can be consolidated into a few, reducing weight, power consumption, and cost. It also becomes safe to add new features through software updates after the car is sold, so that a car's functions can keep improving after purchase.

This requires techniques that let safety-critical functions and other functions run on a single computer without interfering with each other.

Where we are heading

The long-term goal is an open-source software platform on which the timing behavior of each application can be guaranteed from design through operation, even on high-performance general-purpose hardware.

We also intend to apply the results to other fields that require real-time behavior, such as robots, aerospace, and industrial systems, and to release them as open source.

Example thesis topics

The focus is shifting toward computing platforms for software-defined vehicles and toward real-time guarantees for workloads that include AI processing. Students' thesis topics are usually chosen from within these directions, through discussion.

  • Designing and evaluating mixed-criticality service integration on deterministic execution environments such as Lingua Franca
  • Guaranteeing the responsiveness of AI inference on heterogeneous SoCs with NPUs and GPUs
  • Evaluating and improving the real-time behavior of lightweight execution environments: unikernels and containers
  • End-to-end responsiveness through bandwidth guarantees for storage and network I/O
  • Performance analysis and tracing for robot middleware such as ROS 2