g., every relation with
Figure 6. Node contents for PRS and PFRD-H partitioning strategies: (a) PRS
partitioning, and (b) PFRD-H partitioning
(a) (b)
2 6 Furtado
Copyright ?© 2007, Idea Group Inc. Copying or distributing in print or electronic forms without written permission of
Idea Group Inc. is prohibited.
less than 250 MB is to be replicated and those with less than 100 MB are
to be cached into memory for faster access). For our experiments we have
used this simple approach, but we describe a cost model and discuss the
search for optimal partitioning in the next section.
2. Facts: The objective is to find the hash-partitioning attribute that minimizes
repartitioning costs. A reasonable approximation to this objective
is to determine the most frequent equi-join attribute used by the relation.
To do this, the partitioning strategy looks at the frequency of access to
other partitioned relations and chooses the most frequent equi-join attribute
with those relations as the partitioning attribute. We have described
this process in the second section. A more complex approach involves
the search for optimal partitioning, as described in the next section.
By co-locating relation fragments that are frequent equi-join targets, this simple
strategy reduces significantly repartitioning requirements (we have determined experimentally
that WBP achieves an improvent of about 50% over straightforward
primary-key based partitioning (PK) when executing the query of Figure 5 under
the same conditions described later in the experiments).
Pages:
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413