Time-To-Complete Prediction for Data Transfers

Currently, there is no prediction provided to users for the amount of time a particular data transfer from one site in the Worldwide LHC Computing Grid to another will take to complete. To develop a time-to-complete prediction, network performance data and per-file information is gathered from two separate databases and fused, and the resulting cleaned data is fitted using random forest regression. Results are shown for two separate links: the link from CERN Data Centre to Brookhaven National Laboratory's ATLAS data center, and the link from CERN Data Centre to SARA-MATRIX in Amsterdam. A total RMS error of 25.93 minutes between predicted and test data is found for the CERN-PROD -> BNL-ATLAS link, while the CERN-PROD -> SARA-MATRIX link yields a total RMS error of 3.00 minutes.