Fresco: A Public Multi-Institutional Dataset for Understanding HPC System Behavior and Dependability

The scarcity of publicly available operational data from High Performance Computing (HPC) systems hinders research in critical areas like system dependability and resource optimization. While recent efforts, such as the Atlas project, have increased the availability of cluster traces, these datasets often lack fine-grained operational metrics linked to comprehensive job-level attributes across diverse environments. To address this gap, we introduce Fresco, a dataset containing data from 20.9 million jobs spanning 75 months collected from three major academic supercomputing clusters: Purdue’s Anvil and Conte systems, and Texas Advanced Computing Center’s Stampede. Fresco uniquely captures six key performance metrics alongside many job-level attributes such as resource allocations and execution outcomes. We detail our data integration process that transforms and standardizes the heterogeneous data sources into a consistent format. The resulting dataset enables researchers to investigate the relationships between job characteristics, resource consumption patterns, and system performance in academic HPC environments. We make this resource open source at https://www.frescodata.xyz. Our expectation is that this public release will facilitate research and operational improvements that had previously been impossible due to the unavailability of such data.