Databases Reference
In-Depth Information
Figure 6.3
TPCH from October 2006 showing 3TB and 10TB price-performance results
Figure 6.4
Nearly identical performance results for a 0.5 TB and 1.0 TB warehouse.
Each node in the three-node MPP holds one-third of the data. So each node processes a
10
10 join. Although each of the nodes processes this amount, the joins on each
machine occur in parallel, so the execution time is gated by the speed of the slowest
node, namely 10
×
×
10. Had the shared-nothing server been a 30-way complex, then the
computation time would be only one-one-hundredth! This is an idealized scenario, of
course. It assumes that data can be distributed perfectly across nodes, and that the join
data is perfectly collocated. However, in practice even data distribution is common, and
careful selection of the partitioning keys can give good collocation.
