mirror of https://github.com/postgres/postgres
Correct oversight in hashjoin cost estimation: nodeHash sizes its hash
table for an average of NTUP_PER_BUCKET tuples/bucket, but cost_hashjoin was assuming a target load of one tuple/bucket. This was causing a noticeable underestimate of hashjoin costs.
This commit is contained in:
parent
24864d048e
commit
25442d8d2f
|
@ -7,7 +7,7 @@
|
|||
* Portions Copyright (c) 1994, Regents of the University of California
|
||||
*
|
||||
*
|
||||
* $Id: nodeHash.c,v 1.44 2000/01/26 05:56:22 momjian Exp $
|
||||
* $Id: nodeHash.c,v 1.45 2000/04/18 05:43:01 tgl Exp $
|
||||
*
|
||||
*-------------------------------------------------------------------------
|
||||
*/
|
||||
|
@ -221,7 +221,6 @@ ExecEndHash(Hash *node)
|
|||
* create a hashtable in shared memory for hashjoin.
|
||||
* ----------------------------------------------------------------
|
||||
*/
|
||||
#define NTUP_PER_BUCKET 10
|
||||
#define FUDGE_FAC 2.0
|
||||
|
||||
HashJoinTable
|
||||
|
|
|
@ -42,7 +42,7 @@
|
|||
* Portions Copyright (c) 1994, Regents of the University of California
|
||||
*
|
||||
* IDENTIFICATION
|
||||
* $Header: /cvsroot/pgsql/src/backend/optimizer/path/costsize.c,v 1.57 2000/04/12 17:15:19 momjian Exp $
|
||||
* $Header: /cvsroot/pgsql/src/backend/optimizer/path/costsize.c,v 1.58 2000/04/18 05:43:02 tgl Exp $
|
||||
*
|
||||
*-------------------------------------------------------------------------
|
||||
*/
|
||||
|
@ -51,6 +51,7 @@
|
|||
|
||||
#include <math.h>
|
||||
|
||||
#include "executor/nodeHash.h"
|
||||
#include "miscadmin.h"
|
||||
#include "nodes/plannodes.h"
|
||||
#include "optimizer/clauses.h"
|
||||
|
@ -604,12 +605,17 @@ cost_hashjoin(Path *path,
|
|||
run_cost += cpu_operator_cost * outer_path->parent->rows;
|
||||
|
||||
/*
|
||||
* the number of tuple comparisons needed is the number of outer
|
||||
* tuples times the typical hash bucket size, which we estimate
|
||||
* conservatively as the inner disbursion times the inner tuple count.
|
||||
* The number of tuple comparisons needed is the number of outer
|
||||
* tuples times the typical hash bucket size. nodeHash.c tries for
|
||||
* average bucket loading of NTUP_PER_BUCKET, but that goal will
|
||||
* be reached only if data values are uniformly distributed among
|
||||
* the buckets. To be conservative, we scale up the target bucket
|
||||
* size by the number of inner rows times inner disbursion, giving
|
||||
* an estimate of the typical number of duplicates of each value.
|
||||
* We then charge one cpu_operator_cost per tuple comparison.
|
||||
*/
|
||||
run_cost += cpu_operator_cost * outer_path->parent->rows *
|
||||
ceil(inner_path->parent->rows * innerdisbursion);
|
||||
NTUP_PER_BUCKET * ceil(inner_path->parent->rows * innerdisbursion);
|
||||
|
||||
/*
|
||||
* Estimate the number of tuples that get through the hashing filter
|
||||
|
|
|
@ -7,7 +7,7 @@
|
|||
* Portions Copyright (c) 1996-2000, PostgreSQL, Inc
|
||||
* Portions Copyright (c) 1994, Regents of the University of California
|
||||
*
|
||||
* $Id: nodeHash.h,v 1.15 2000/01/26 05:58:05 momjian Exp $
|
||||
* $Id: nodeHash.h,v 1.16 2000/04/18 05:43:00 tgl Exp $
|
||||
*
|
||||
*-------------------------------------------------------------------------
|
||||
*/
|
||||
|
@ -16,6 +16,9 @@
|
|||
|
||||
#include "nodes/plannodes.h"
|
||||
|
||||
/* NTUP_PER_BUCKET is exported because planner wants to see it */
|
||||
#define NTUP_PER_BUCKET 10
|
||||
|
||||
extern TupleTableSlot *ExecHash(Hash *node);
|
||||
extern bool ExecInitHash(Hash *node, EState *estate, Plan *parent);
|
||||
extern int ExecCountSlotsHash(Hash *node);
|
||||
|
|
Loading…
Reference in New Issue