Speaker
Timothy Adam Barrass
(University of Bristol)
Description
Distributed data management at LHC scales is a staggering task, accompanied by
equally challenging practical management issues with storage systems and wide-area
networks. CMS data transfer management system, PhEDEx, is designed to handle this
task with minimum operator effort, automating the workflows from large scale
distribution of HEP experiment datasets down to reliable and scalable transfers of
individual files over frequently unreliable infrastructure. PhEDEx has been designed
and proven to scale beyond the current CMS needs. Few of the techniques we have used
are novel, but rarely documented in HEP. We describe many of the techniques we have
used to make the system robust and able to deliver high performance. On schema and
data organisation we describe our use of hierarchical data organisation, separation
of active and inactive data, and tuning the database for the data and access
patterns. Regarding monitoring we describe our use of optimised queries, moving
queries away from hot tables, and using multi-level performance histograms to
precalculate partial aggregated results. Robustness applies to both detecting and
recovering from local errors, and robustness in the distributed environment. We
describe the coding patterns we use for error-resilient and selfhealing agents for
the former, and the breakdown of handshakes in file transfer, routing files to
destinations, and in managing site presence for the latter.
Primary authors
Lassi Tuura
(Northeastern University)
Timothy Adam Barrass
(University of Bristol)
