HOPPERFISH: Holistic Profiling with Portable Extensible and Robust Framework Intended for Systems with Heterogeneity

We introduce HOPPERFISH, a holistic profiling framework that unifies analysis across the application, runtime, microarchitecture, and hardware layers to streamline robust feature correlation in heterogeneous computing systems. HOPPERFISH provides comprehensive insights into dynamic workloads, hardware configurations, and scheduling policies by capturing features across the system stack for any heterogeneous System on Chip (SoC), whether it is an off-the-shelf platform or an architecture emulated on an FPGA. The framework enables data-driven analysis for real-world applications and unsupervised learning for heterogeneous systems where features vary dynamically and cannot be labeled in advance. As a use case, we utilize HOPPERFISH in the anomaly detection context for heterogeneous systems and build an autoencoder model under varying workload, hardware, and scheduling scenarios without the need to retrain separate models for each scenario. HOPPERFISH extracts correlations among features across the entire system stack, which leads to requiring a smaller number of parameters to build a representative model for anomaly detection. This compressed form leads to the implementation of a robust and accurate yet lightweight model that can detect abnormal behaviors in real-time based on unsupervised behavior analysis. As a proof of concept, we also demonstrate hardware deployment of the real-time anomaly detection model with a latency of 18 µs on MPSoC ZCU102 FPGA consuming 6.3 µJ, achieving 16.67× lower latency and 117× less energy consumption compared to its software implementation. HOPPERFISH is open-source and accompanied by an extensible workload benchmark suite, facilitating broader research tasks for the development of future heterogeneous computing systems, including security analysis, fault diagnosis, and data-driven design optimization decisions.