Description
The ATLAS experiment uses a tiered data Grid architecture that enables
possibly overlapping subsets, or replicas, of the original set to be
located across the ATLAS collaboration. The full set of experiment
data is located at a single Tier 0 site, and then subsets of the data
are located at national Tier 1 sites, smaller subsets at smaller
regional Tier 2 sites, and so on. In order to understand the data
needs, both in terms of access, replication policy, and storage
capacity, we need good estimations of resource needs for data
manipulation. Specifically, we envision a time when a user will want
to determine which is more expedient, downloading a replica from a
site or recreating it from scratch.
This paper presents our technique to predict the behavior of ATLAS
applications, and then to combine this information with Internet link
bandwidth estimation to improve resource usage in the ATLAS Grid
environment. We studied the parameters that affect the execution time
performance of event generation, detector simulation, and event
reconstruction. Our results show that we can achieve predictions
within 10-40% of the execution time (depending on the application),
better than many other pragmatic prediction techniques. We implemented
a software package to provide data transfer bandwidth estimation and
execution time prediction that can be used with the Chimera software
to aid in managing application execution and to improve resource usage
for ATLAS.
