Accelerating HPC Performance Modeling Via Diffusion-Based Synthetic Data Generation

Abstract

Machine learning-based performance modeling is increasingly used to guide compiler optimizations and autotuning of performance-critical applications. However, constructing largescale, representative training datasets remains a fundamental challenge due to the high cost of empirical profiling and limited generalizability across architectures. In this paper, we present DIMED, a diffusion-based framework for generating scalable, high-fidelity synthetic performance datasets. Our approach supports multiple modalities, including static code features, compiler graph representations, and hardware performance counters capturing runtime behavior. A key innovation is a dictionaryencoded representation that enables effective modeling of multilabel outputs and long-tailed distributions commonly observed in performance data. Experimental results demonstrate that our method produces synthetic data with high statistical fidelity, achieving an average Kolmogorov-Smirnov similarity score of 97% across four representative performance datasets. We evaluate the utility of the generated data in four downstream optimization tasks: (i) workload classification, (ii) power-consumption classification, (iii) compiler optimization sequence prediction, and (iv) job-state classification. Models trained with synthetic data achieve accuracy exceeding those trained on real data alone, providing up to 13% improvement in classification accuracy. More importantly, when integrated into optimization workflows, our approach maintains competitive optimization effectiveness relative to state-of-the-art methods while significantly reducing dependence on costly large-scale data collection.