the n?? of nodes for query Qa.
The performance of replica-based strategies (especially PRS) is much worse than
partitioning-based ones (WBP, WBP+JB), because nodes have to process large
replicated datasets. Additionally, (WBP+JB) improves response time further, as
early-selection functionality reduces the amount of data that must be processed
and repartitioned. Of course bitmap join indexes must be available and their usefulness
depends on the selectivity of query select conditions. On the other hand, if the
220 Furtado
Copyright ?© 2007, Idea Group Inc. Copying or distributing in print or electronic forms without written permission of
Idea Group Inc. is prohibited.
available network bandwidth is low, strategies using replicas (e.g., PRS, PFRD-H)
can exhibit better performance relative to those relying on partitioning (WBP), as
shown in Figure 10a. Still, early-selection (WBP+JB) was the best strategy because
it is not very dependent on repartitioning.
We also subjected our simulator to conformance tests, to evaluate whether its simulation
is sufficiently accurate for our purposes. Figure 11 shows a result from those
tests. We ran WBP and PRS on a system with the characteristics: TPC-H 25GB;
commercial DBMS; each node with 3 GHz Intel Pentium 4 processor; 1 GB memory;
200 GB SATA II disk; 1 GB/s network; query Qa). Cost-based optimization was used,
the schema objects were analyzed and the best query plan suggestion was chosen
(the default execution plan had very bad performance for PRS).
Pages:
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420