An Empirical Study of Machine Learning-Based Synthetic Job Trace Generation Methods

Job traces play an increasingly important role in today’s data-driven and machine learning-based computing system optimizations. However, the lack of sufficient public real-world job traces and the difficulty of publishing more job traces from large computing centers make generating synthetic job traces necessary and important. In this study, we conduct a comparative study of five different machine learning-based data generation methods (i.e., GAN, CTGAN, TVAE, Gaussian Copula, Copula GAN) on three different performance metrics (i.e., CDF plots, statistical metrics, simulation-based scheduling). We also conducted a comparison with the state-of-the-art manual statistical job trace generation method, which was the default method before the new generative models were recently proposed. Our results suggest that many of the machine learning-based data generation methods can consistently produce high-quality job traces without significant manual fine-tuning efforts. The generated trace can be comparable to or even better than the trace generated from highly-optimized manual statistical models. We believe such results indicate the possibility of building an effective pipeline to generate high-fidelity job traces given an arbitrary real-world job trace, allowing for quick and large-scale release of job traces. We provide all the generated traces at https://​github.​com/​DIR-LAB/​Gen-Parallel-Workloads for others to use.