By enhancing our Bricks Grid simulator to be able to simulate Data
Grid scenarios, we compare the performance of different Data Grid
models in the Grid Datafarm architecture; the architecture is mainly
categorized into the central and the tier models but with varying
scheduling and replication strategies, under realistic assumptions of
job processing for the CERN LHC experiments. The Grid Datafarm
architecture provides a global parallel striping file system with
on-line petascale storage using a grid of clusters to achieve scalable
disk I/O bandwidth. We propose simple on-line scheduling and
replication policies for tier-based Data Grid models and investigate
the performance of central and tier models with the combinations of
our scheduling and replication policies. Our simulations show the
central model performance improves with higher Tier 0 site
performance, the system quick became unstable for resource-starved
situations. For the tier model, in comparing its aggregate performance
to the equivalent amount of resources for the central model, it is
beaten hands down. However, if we have greater amount of resources we
can sustain system stability and achieve higher performance while each
tier being smaller than the central model. In such a case employing
speculative class of scheduling and replication policies prove to be
effective.