QoS-Oriented Grid-Enabled Data Warehouses - Data Warehousing Design and Advanced Engineering Applications - page 163

Figure 5. Experimental setup description

local caching. But chunks of results or computed
chunks can also be considered as candidates for
replication by Replication Manager. In such con-
text, when a local scheduler evaluates that a certain
missing chunk would enable the site to execute a
task by a certain deadline that would be achieved
without such chunk, the local scheduler must send
to the Replication Manager the identification of
such chunk. Then, the RM would consider the
chunk as a dataset candidate for replication in
its benefit-based dynamic replica selection and
placement mechanism.

sizes are generated considering Pareto's power-law
distribution [which fits well for data grids files
(Sulistio et al, 2007)], with a 1Gb mean fragment
size. Initially, each site stores the same number of
facts table fragments (11 fragments).
Two distinct network topologies are used in our
tests (represented in Figure 6). The first one is a
hierarchical topology, in which sites are organized
in a binary tree according to their ids [hierarchical
topologies are also considered in real projects, like
the LHC Computing Grid Project - data storage
and analysis project for CERN's Large Hadron
Collider (Bird et al, 2005)]. The second network
model is a nonhierarchical topology, dual ring
topology. In such topology, there is no central
root site (eliminating a possible bottleneck in the
system) and each site is directly connected to two
other sites. Data movement in the rings is done in
opposite directions. In all the tests, we consider
a data transfer rate of 50Mbps and a latency of
10 milliseconds.
The considered query workload is composed
of 1,000 tasks (re-written queries). Tasks' sizes
vary about 2,000 kMIPS ± 30%, which means
that a typical task execution would take about 30
minutes in the least powerful site and 40 seconds

ExPERIMENTAL EVALUATION

The QoS-oriented scheduling and dynamic repli-
cation mechanisms were experimentally evaluated
in a simulation environment. The experimental
setup is composed by 11 sites, which were inspired
in the experimental testbed used in (Sulistio et al,
2007) and in the LHC Computing Grid Project
sites (Bird et al, 2005). Figure 5 presents the main
characteristics of the used sites.
A Star Schema-based DW is considered. Facts
table is partitioned into 121 fragments. Fragments