Holistic framework that analyzes cross-layer behavior in SoCs to improve reliability and efficiency.

What is it about?

Modern computer chips bring together many kinds of processors, such as CPUs, GPUs, and custom accelerators, to deliver high performance while saving energy. These systems are powerful, but their complexity makes them difficult to understand and manage. Each application may run differently depending on the hardware resources it uses and the scheduling decisions made at runtime, and existing tools usually look only at one layer of the system rather than the full picture. In this work we introduce HOPPERFISH, an open source framework that provides a holistic way to monitor and analyze these heterogeneous systems. HOPPERFISH collects detailed information across applications, runtime software, microarchitecture, and hardware, and then combines these features to reveal how the system behaves under different workloads. We show that this holistic profiling can be used to build lightweight machine learning models that detect unusual or harmful behavior in real time, without the need for manually labeled data. Our experiments demonstrate that the same model can be applied across different platforms including CPUs, GPUs, and FPGA based systems, and that a hardware implementation of the model achieves extremely fast response with very low energy consumption. This makes it possible to detect anomalies such as security threats or performance issues while the system is running, rather than only during post-analysis. By improving visibility across all layers of the computing stack, HOPPERFISH enables researchers and engineers to design more reliable, secure, and energy efficient systems for demanding applications ranging from autonomous vehicles and communications to industrial and defense technologies.

Featured Image

Photo by Igor Omilaev on Unsplash

Photo by Igor Omilaev on Unsplash

Why is it important?

HOPPERFISH is important because it addresses a growing gap in how we study and manage modern computer systems. As chips integrate more types of processors and accelerators, their complexity increases faster than the tools available to monitor them. Traditional profiling methods often fail to capture how decisions made at the runtime layer interact with microarchitecture and hardware behavior. Our work is the first to provide a single framework that unifies these perspectives and makes them accessible for both analysis and machine learning. This is timely because real-world systems now demand reliable, low-overhead anomaly detection to ensure security and efficiency. Our approach shows that it is possible to achieve this without sacrificing speed or energy, even on constrained hardware like FPGAs. By demonstrating portability across different platforms and delivering practical hardware implementations, HOPPERFISH lays the groundwork for future systems that can adapt intelligently and remain secure in the face of increasing complexity.