Latest results of validation studies (workload run repeated 20 times):
- https://twiki.cern.ch/twiki/bin/view/HEPIX/HEPWorkloadReproducibility
- ATLAS:
- Spread in the order of 10...20% when processing 5 events
- Improved by increasing number of events, but variations in the scores of repeated runs are still about 10% (500 events, runtime: ~0.5 h), and still greater than 5% even when processing 2500 events (runtime: ~2.5 h)
- How to improve?
- CMS:
- Spread of around 2% when processing 20 events (runtime: ~15')
- CPU_score as well as througput_score
- Variations of the order of 10...20% in first runs on another host, still under investigation
- Few high-scoring jobs found
- Output folders contain a file named step1.root which is about 325 MB (in case of 100 events), but one step1.root file of the higher scoring jobs is only 125 MB
- No critical error messages found in the corresponding output logfiles
- Looking for volunteers to run the workload at other sites for double-checking