I used to think the cost of an NLJOIN between an outer and an inner
subquery should be at least

card(outer) * cost(inner),

but the optimizer of DB2/NT64 9.7.5 at standard settings surprises me
with the following:

2.60477e+008
NLJOIN
( 2)
350361
42426.5
/----+-----\
3.86338e+006 67.422
TBSCAN IXSCAN
( 3) ( 6)
35725.5 25.7914
15576.4 2

According to my understanding, the resulting cost(2) (= cost of node 2)
should be at least card(3) * cost(6), which is 3.86338e+006 * 25.7914 =
99.6e6; that is, 100 million instead of the measly 350361 it comes up with.

The actual performance of the plan is as terrible as it looks (I aborted
it after several hours) and confirms that 100 million is closer to the
truth than 350k. Better plans would presumably be available, but do not
get chosen as long as this plan is calculated to be so cheap.

I have seen such strange cost estimates before in the context of more
complex queries, but this is the first time I have seen it in a simple
two-table join.

Can anyone explain what is wrong here?

Regards,

Oliver Schoett

On 06/20/2012 07:01 PM, Oliver Schoett wrote:
If the plan estimate does not agree with reality, I'd first check if statistcs are OK.
/dg

danfan46 wrote:

Good point - I had only inspected the statistics visually. So I created
them afresh with

RUNSTATS ON TABLE ...
on all columns
WITH DISTRIBUTION
AND DETAILED INDEXES ALL
ALLOW WRITE ACCESS

on both the tables involved. The access plan has not changed:

2.60477e+008
NLJOIN
( 2)
350361
42426.5
/----+-----\
3.86338e+006 67.422
TBSCAN IXSCAN
( 3) ( 6)
35725.5 25.7914
15576.4 2

Interestingly, adding "for read only optimize for 1 row" at the end of
the select statement changes not the NLJOIN but its cost estimate:

2.60477e+008
NLJOIN
( 2)
757934
67353.3
/-----+------\
3.86338e+006 67.422
TBSCAN IXSCAN
( 3) ( 4)
63038 25.7914
34066 2

The cost change of node(3) is OK, as a different index is used to obtain
the set of join keys (wider, hence requiring more I/O, but not requiring
a sort, hence delivering the first row more quickly). This should
increase the total cost of the NLJOIN accordingly (by about 30K), as the
result set of node(3) is the same as in the original plan (it is the
result of a full index scan as there are no conditions except the join
conditions in the query).

However, the cost estimate of the NLJOIN is now ~760k instead of the
correct ~100 million, so just as wrong, only different. This does not
make any sense to me whatsoever.

Regards,

Oliver Schoett