Raw Experimental Results for GPU Storage Resilience Assessment with SARA, GEREM Storage-EFM, and Fault Injection

It contains CSV files with application-level and component-level outcome counts, outcome rates, and execution-time measurements. Outcome fields include Masked, SDC, DUE, and Unknown where applicable. Count fields represent modeled fault-site counts or sampled fault outcomes depending on the method.

Timing fields are reported in seconds.

Summary

Markdown files and PDF figures are included under the paper-result summary directory inside the dataset.The dataset contains 327 CSV files, 7 Markdown summary files, 3 PDF figures, and 1 JSON manifest. For each architecture, SARA contains 28 CSV files with 308 data rows, each GEREM sampling configuration contains 28 CSV files with 308 data rows, and FI contains 51 CSV files with 51 data rows. Some application-component combinations are absent because the corresponding benchmark does not exercise that storage component; these omissions are intentional and are not missing data.This is a computational experiment dataset rather than geographic or temporal observational data; therefore, spatial coverage, spatial resolution, and temporal resolution are not applicable.

The main experimental dimensions are GPU architecture, benchmark application, storage component, and resilience assessment method. Sampling-based methods may include sampling uncertainty, while deterministic analysis results depend on the modeled traces, storage fault model, and exact-output oracle.All files use common formats: CSV for raw numerical results, Markdown for summary tables, JSON for metadata, and PDF for generated figures.

They can be opened with standard spreadsheet software, text editors, JSON viewers, and PDF readers.