Database Reference
In-Depth Information
Speed up of Parallel FFT with Logical Window Size 8192
10
9
Ideal speed-up
Window split
Window distribute
8
7
6
5
4
3
2
1
1
2
3
4
5
6
7
8
9
10
Number of Processing Nodes
Figure 11.8
FFT speedup for parallelism of degree four with different dis-
tribution templates.
strategies have advantages in specific situations. If the continuous query is
to be executed with a limited number of compute nodes, so that the load of
the compute nodes exceeds the load of the partition and combine phases, the
window split is preferable since it utilizes query semantics to achieve a more
scalable parallel execution. However, if the system has resources allowing a
high degree of parallelism where partition and combine nodes become more
loaded than the compute nodes, window distribute may have better perfor-
mance depending on the cost of partitioning and combining SQFs. This is
further illustrated by Figure 11.8, which shows the speed-up of the different
partitioning schemes. The window split template provides better speed-up
when the number of processors is limited. The tree-shaped window distribute
provides the best performance in Figure 11.7. However, it also uses the most
compute nodes (8) and does not provide as good speed-up.
11.3.2 Stream Processes in SCSQ
SCSQ enables customized parallelization of continuous queries defined com-
pletely in the query language SCSQL (pronounced sis-kel). In SCSQL, parallel
computations in queries and views are specified in terms of stream processes
(SPs) that are first-class objects in CQs. The CQs can call process construc-
tion functions that execute stream subqueries assigned to some CPU. Such
queries can be used to define query functions that parallelize computations
over streams. The user can specify massively parallel stream computations by
defining sets of SPs executing arbitrary subqueries. Properties of the different
