This repository contains the Python code and Jupyter notebooks that were used to perform the analyses and produce the plots for the manuscript Application Failures and Machine Computational Efficiency, by Graziani, Lusch, and Messer, submitted to the SC-2025 conference.
The utility code is contained in the file efficiency.py.
The file jobs2024.csv contains a database of 331,640 production jobs from the Frontier supercomputer at ORNL. These data are one year (CY 2024) of jobs from the INCITE and ALCC programs, with a further selection for jobs of more than 1 node in size, to exclude interactive and debug jobs.
The notebook Machine_Efficiency_Analysis.ipynb imports the functions in the efficiency.py file to compute machine efficiency assuming the computational load that characterizes 1 year of production jobs on Frontier. The analysis results in the paper, and all but one of the figures, are output by this notebook.
The notebook Optimal_Checkpointing.ipynb contains code for computing the optimal checkpointing cadence, essentially updating the work of Daly (2006) to give cadences in usage (node-hrs) rather than in time (hrs). This notebook also produces Figure 1 of the paper.