Examining Job Exit Status Reveals Surprising Information about Computer Operations

Investigators: A Sim, W Yoo

Source Data

NERSC genepool cluster batch queue job exit status plus memory usage, CPU usage and so on

What We Found

• Clustering reveals strange outliers as red circles on the right. Outliers can detect the performance anomalies.
• Job exit status could be predicted with 99% accuracy
• Dips in exit status prediction rate coincide with known system-wide failures

What's Next

• Performance prediction (time, I/O, resource consumption) by job characterization and modeling
• Online failure detection using runtime measurements for early actions on jobs expected to fail