Skip to content

Commit e5253fd

Browse files
robertmhaasrobertmhaas
committed
Add parallel_leader_participation GUC.
Sometimes, for testing, it's useful to have the leader do nothing but read tuples from workers; and it's possible that could work out better even in production. Thomas Munro, reviewed by Amit Kapila and by me. A few final tweaks by me. Discussion: http://postgr.es/m/CAEepm=2U++Lp3bNTv2Bv_kkr5NE2pOyHhxU=G0YTa4ZhSYhHiw@mail.gmail.com
1 parent 7518049 commit e5253fd

File tree

10 files changed

+205
-9
lines changed

10 files changed

+205
-9
lines changed

‎doc/src/sgml/config.sgml

Lines changed: 26 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -4265,6 +4265,32 @@ SELECT * FROM parent WHERE key = 2400;
42654265
</listitem>
42664266
</varlistentry>
42674267

4268+
<varlistentry id="guc-parallel-leader-participation" xreflabel="parallel_leader_participation">
4269+
<term>
4270+
<varname>parallel_leader_participation</varname> (<type>boolean</type>)
4271+
<indexterm>
4272+
<primary>
4273+
<varname>parallel_leader_participation</varname> configuration
4274+
parameter
4275+
</primary>
4276+
</indexterm>
4277+
</term>
4278+
<listitem>
4279+
<para>
4280+
Allows the leader process to execute the query plan under
4281+
<literal>Gather</literal> and <literal>Gather Merge</literal> nodes
4282+
instead of waiting for worker processes. The default is
4283+
<literal>on</literal>. Setting this value to <literal>off</literal>
4284+
reduces the likelihood that workers will become blocked because the
4285+
leader is not reading tuples fast enough, but requires the leader
4286+
process to wait for worker processes to start up before the first
4287+
tuples can be produced. The degree to which the leader can help or
4288+
hinder performance depends on the plan type, number of workers and
4289+
query duration.
4290+
</para>
4291+
</listitem>
4292+
</varlistentry>
4293+
42684294
<varlistentry id="guc-force-parallel-mode" xreflabel="force_parallel_mode">
42694295
<term><varname>force_parallel_mode</varname> (<type>enum</type>)
42704296
<indexterm>

‎src/backend/executor/nodeGather.c

Lines changed: 5 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -38,6 +38,7 @@
3838
#include "executor/nodeSubplan.h"
3939
#include "executor/tqueue.h"
4040
#include "miscadmin.h"
41+
#include "optimizer/planmain.h"
4142
#include "pgstat.h"
4243
#include "utils/memutils.h"
4344
#include "utils/rel.h"
@@ -73,7 +74,8 @@ ExecInitGather(Gather *node, EState *estate, int eflags)
7374
gatherstate->ps.ExecProcNode = ExecGather;
7475

7576
gatherstate->initialized = false;
76-
gatherstate->need_to_scan_locally = !node->single_copy;
77+
gatherstate->need_to_scan_locally =
78+
!node->single_copy && parallel_leader_participation;
7779
gatherstate->tuples_needed = -1;
7880

7981
/*
@@ -193,9 +195,9 @@ ExecGather(PlanState *pstate)
193195
node->nextreader = 0;
194196
}
195197

196-
/* Run plan locally if no workers or not single-copy. */
198+
/* Run plan locally if no workers or enabled and not single-copy. */
197199
node->need_to_scan_locally = (node->nreaders == 0)
198-
|| !gather->single_copy;
200+
|| (!gather->single_copy && parallel_leader_participation);
199201
node->initialized = true;
200202
}
201203

‎src/backend/executor/nodeGatherMerge.c

Lines changed: 4 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -23,6 +23,7 @@
2323
#include "executor/tqueue.h"
2424
#include "lib/binaryheap.h"
2525
#include "miscadmin.h"
26+
#include "optimizer/planmain.h"
2627
#include "utils/memutils.h"
2728
#include "utils/rel.h"
2829

@@ -233,8 +234,9 @@ ExecGatherMerge(PlanState *pstate)
233234
}
234235
}
235236

236-
/* always allow leader to participate */
237-
node->need_to_scan_locally = true;
237+
/* allow leader to participate if enabled or no choice */
238+
if (parallel_leader_participation || node->nreaders == 0)
239+
node->need_to_scan_locally = true;
238240
node->initialized = true;
239241
}
240242

‎src/backend/optimizer/path/costsize.c

Lines changed: 8 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -5137,7 +5137,6 @@ static double
51375137
get_parallel_divisor(Path *path)
51385138
{
51395139
double parallel_divisor = path->parallel_workers;
5140-
double leader_contribution;
51415140

51425141
/*
51435142
* Early experience with parallel query suggests that when there is only
@@ -5150,9 +5149,14 @@ get_parallel_divisor(Path *path)
51505149
* its time servicing each worker, and the remainder executing the
51515150
* parallel plan.
51525151
*/
5153-
leader_contribution = 1.0 - (0.3 * path->parallel_workers);
5154-
if (leader_contribution > 0)
5155-
parallel_divisor += leader_contribution;
5152+
if (parallel_leader_participation)
5153+
{
5154+
double leader_contribution;
5155+
5156+
leader_contribution = 1.0 - (0.3 * path->parallel_workers);
5157+
if (leader_contribution > 0)
5158+
parallel_divisor += leader_contribution;
5159+
}
51565160

51575161
return parallel_divisor;
51585162
}

‎src/backend/optimizer/plan/planner.c

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -61,6 +61,7 @@
6161
/* GUC parameters */
6262
double cursor_tuple_fraction = DEFAULT_CURSOR_TUPLE_FRACTION;
6363
int force_parallel_mode = FORCE_PARALLEL_OFF;
64+
bool parallel_leader_participation = true;
6465

6566
/* Hook for plugins to get control in planner() */
6667
planner_hook_type planner_hook = NULL;

‎src/backend/utils/misc/guc.c

Lines changed: 10 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1676,6 +1676,16 @@ static struct config_bool ConfigureNamesBool[] =
16761676
NULL, NULL, NULL
16771677
},
16781678

1679+
{
1680+
{"parallel_leader_participation", PGC_USERSET, RESOURCES_ASYNCHRONOUS,
1681+
gettext_noop("Controls whether Gather and Gather Merge also run subplans."),
1682+
gettext_noop("Should gather nodes also run subplans, or just gather tuples?")
1683+
},
1684+
&parallel_leader_participation,
1685+
true,
1686+
NULL, NULL, NULL
1687+
},
1688+
16791689
/* End-of-list marker */
16801690
{
16811691
{NULL, 0, 0, NULL, NULL}, NULL, false, NULL, NULL, NULL

‎src/backend/utils/misc/postgresql.conf.sample

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -163,6 +163,7 @@
163163
#effective_io_concurrency = 1 # 1-1000; 0 disables prefetching
164164
#max_worker_processes = 8 # (change requires restart)
165165
#max_parallel_workers_per_gather = 2 # taken from max_parallel_workers
166+
#parallel_leader_particulation = on
166167
#max_parallel_workers = 8 # maximum number of max_worker_processes that
167168
# can be used in parallel queries
168169
#old_snapshot_threshold = -1 # 1min-60d; -1 disables; 0 is immediate

‎src/include/optimizer/planmain.h

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -29,6 +29,7 @@ typedef enum
2929
#define DEFAULT_CURSOR_TUPLE_FRACTION 0.1
3030
extern double cursor_tuple_fraction;
3131
extern int force_parallel_mode;
32+
extern bool parallel_leader_participation;
3233

3334
/* query_planner callback to compute query_pathkeys */
3435
typedef void (*query_pathkeys_callback) (PlannerInfo *root, void *extra);

‎src/test/regress/expected/select_parallel.out

Lines changed: 113 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -34,6 +34,49 @@ select count(*) from a_star;
3434
50
3535
(1 row)
3636

37+
-- test with leader participation disabled
38+
set parallel_leader_participation = off;
39+
explain (costs off)
40+
select count(*) from tenk1 where stringu1 = 'GRAAAA';
41+
QUERY PLAN
42+
---------------------------------------------------------
43+
Finalize Aggregate
44+
-> Gather
45+
Workers Planned: 4
46+
-> Partial Aggregate
47+
-> Parallel Seq Scan on tenk1
48+
Filter: (stringu1 = 'GRAAAA'::name)
49+
(6 rows)
50+
51+
select count(*) from tenk1 where stringu1 = 'GRAAAA';
52+
count
53+
-------
54+
15
55+
(1 row)
56+
57+
-- test with leader participation disabled, but no workers available (so
58+
-- the leader will have to run the plan despite the setting)
59+
set max_parallel_workers = 0;
60+
explain (costs off)
61+
select count(*) from tenk1 where stringu1 = 'GRAAAA';
62+
QUERY PLAN
63+
---------------------------------------------------------
64+
Finalize Aggregate
65+
-> Gather
66+
Workers Planned: 4
67+
-> Partial Aggregate
68+
-> Parallel Seq Scan on tenk1
69+
Filter: (stringu1 = 'GRAAAA'::name)
70+
(6 rows)
71+
72+
select count(*) from tenk1 where stringu1 = 'GRAAAA';
73+
count
74+
-------
75+
15
76+
(1 row)
77+
78+
reset max_parallel_workers;
79+
reset parallel_leader_participation;
3780
-- test that parallel_restricted function doesn't run in worker
3881
alter table tenk1 set (parallel_workers = 4);
3982
explain (verbose, costs off)
@@ -400,6 +443,49 @@ explain (costs off, verbose)
400443
(11 rows)
401444

402445
drop function simple_func(integer);
446+
-- test gather merge with parallel leader participation disabled
447+
set parallel_leader_participation = off;
448+
explain (costs off)
449+
select count(*) from tenk1 group by twenty;
450+
QUERY PLAN
451+
----------------------------------------------------
452+
Finalize GroupAggregate
453+
Group Key: twenty
454+
-> Gather Merge
455+
Workers Planned: 4
456+
-> Partial GroupAggregate
457+
Group Key: twenty
458+
-> Sort
459+
Sort Key: twenty
460+
-> Parallel Seq Scan on tenk1
461+
(9 rows)
462+
463+
select count(*) from tenk1 group by twenty;
464+
count
465+
-------
466+
500
467+
500
468+
500
469+
500
470+
500
471+
500
472+
500
473+
500
474+
500
475+
500
476+
500
477+
500
478+
500
479+
500
480+
500
481+
500
482+
500
483+
500
484+
500
485+
500
486+
(20 rows)
487+
488+
reset parallel_leader_participation;
403489
--test rescan behavior of gather merge
404490
set enable_material = false;
405491
explain (costs off)
@@ -508,6 +594,33 @@ select string4 from tenk1 order by string4 limit 5;
508594
AAAAxx
509595
(5 rows)
510596

597+
-- gather merge test with 0 workers, with parallel leader
598+
-- participation disabled (the leader will have to run the plan
599+
-- despite the setting)
600+
set parallel_leader_participation = off;
601+
explain (costs off)
602+
select string4 from tenk1 order by string4 limit 5;
603+
QUERY PLAN
604+
----------------------------------------------
605+
Limit
606+
-> Gather Merge
607+
Workers Planned: 4
608+
-> Sort
609+
Sort Key: string4
610+
-> Parallel Seq Scan on tenk1
611+
(6 rows)
612+
613+
select string4 from tenk1 order by string4 limit 5;
614+
string4
615+
---------
616+
AAAAxx
617+
AAAAxx
618+
AAAAxx
619+
AAAAxx
620+
AAAAxx
621+
(5 rows)
622+
623+
reset parallel_leader_participation;
511624
reset max_parallel_workers;
512625
SAVEPOINT settings;
513626
SET LOCAL force_parallel_mode = 1;

‎src/test/regress/sql/select_parallel.sql

Lines changed: 36 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -19,6 +19,22 @@ explain (costs off)
1919
select count(*) from a_star;
2020
select count(*) from a_star;
2121

22+
-- test with leader participation disabled
23+
set parallel_leader_participation = off;
24+
explain (costs off)
25+
select count(*) from tenk1 where stringu1 = 'GRAAAA';
26+
select count(*) from tenk1 where stringu1 = 'GRAAAA';
27+
28+
-- test with leader participation disabled, but no workers available (so
29+
-- the leader will have to run the plan despite the setting)
30+
set max_parallel_workers = 0;
31+
explain (costs off)
32+
select count(*) from tenk1 where stringu1 = 'GRAAAA';
33+
select count(*) from tenk1 where stringu1 = 'GRAAAA';
34+
35+
reset max_parallel_workers;
36+
reset parallel_leader_participation;
37+
2238
-- test that parallel_restricted function doesn't run in worker
2339
alter table tenk1 set (parallel_workers = 4);
2440
explain (verbose, costs off)
@@ -157,6 +173,16 @@ explain (costs off, verbose)
157173

158174
drop function simple_func(integer);
159175

176+
-- test gather merge with parallel leader participation disabled
177+
set parallel_leader_participation = off;
178+
179+
explain (costs off)
180+
select count(*) from tenk1 group by twenty;
181+
182+
select count(*) from tenk1 group by twenty;
183+
184+
reset parallel_leader_participation;
185+
160186
--test rescan behavior of gather merge
161187
set enable_material = false;
162188

@@ -192,6 +218,16 @@ set max_parallel_workers = 0;
192218
explain (costs off)
193219
select string4 from tenk1 order by string4 limit 5;
194220
select string4 from tenk1 order by string4 limit 5;
221+
222+
-- gather merge test with 0 workers, with parallel leader
223+
-- participation disabled (the leader will have to run the plan
224+
-- despite the setting)
225+
set parallel_leader_participation = off;
226+
explain (costs off)
227+
select string4 from tenk1 order by string4 limit 5;
228+
select string4 from tenk1 order by string4 limit 5;
229+
230+
reset parallel_leader_participation;
195231
reset max_parallel_workers;
196232

197233
SAVEPOINT settings;

0 commit comments

Comments
 (0)