Обсуждение: Forcing using index instead of sequential scan?
I have been testing the performance of PostgreSQL using the simple tool found at http://benchw.sourceforge.net however I have found that all the queries it run execute with sequential scans. The website where the code runs has examples of the execution plan using indexes.
When I disable the sequential plan query 0 and query 1 run faster ( http://benchw.sourceforge.net/benchw_results_postgres_history.html ) by using the indexes as suggested by the website.
I have tried increasing the effective_cache_size and reducing the random_page_cost to try and force the optimiser to use the index but it always uses the sequential scan.
What is the best way to force the use of indexes in these queries? Currently testing with version 8.1.4.
Regards
Robin Smith
British Telecommunications plc Registered office: 81 Newgate Street London EC1A 7AJ
Registered in England no. 1800000
This electronic message contains information from British Telecommunications plc which may be privileged and confidential. The information is intended to be for the use of the individual(s) or entity named above. If you are not the intended recipient, be aware that any disclosure, copying, distribution or use of the contents of this information is prohibited. If you have received this electronic message in error, please notify us by telephone or e-mail (to the number or address above) immediately.
More information from the query:-
explain analyze
SELECT
d0.dmth,
count(f.fval )
FROM
dim0 AS d0,
fact0 AS f
WHERE d0.d0key = f.d0key
AND d0.ddate BETWEEN '2010-01-01' AND '2010-12-28'
GROUP BY
d0.dmth
;
QUERY PLAN
------------------------------------------------------------------------
-------------------------------------------------------------
HashAggregate (cost=336998.83..336998.84 rows=1 width=8) (actual
time=33823.124..33823.134 rows=12 loops=1)
-> Hash Join (cost=214.83..335343.83 rows=331000 width=8) (actual
time=61.065..33605.343 rows=336000 loops=1)
Hash Cond: ("outer".d0key = "inner".d0key)
-> Seq Scan on fact0 f (cost=0.00..281819.00 rows=10000000
width=8) (actual time=12.766..28945.036 rows=10000000 loops=1)
-> Hash (cost=214.00..214.00 rows=331 width=8) (actual
time=31.120..31.120 rows=336 loops=1)
-> Seq Scan on dim0 d0 (cost=0.00..214.00 rows=331
width=8) (actual time=26.362..30.895 rows=336 loops=1)
Filter: ((ddate >= '2010-01-01'::date) AND (ddate
<= '2010-12-28'::date))
Total runtime: 33823.220 ms
(8 rows)
benchw=# \d fact0
Table "public.fact0"
Column | Type | Modifiers
--------+------------------------+-----------
d0key | integer | not null
d1key | integer | not null
d2key | integer | not null
fval | integer | not null
ffill | character varying(100) | not null
Indexes:
"fact0_d0key" btree (d0key)
"fact0_d1key" btree (d1key)
"fact0_d2key" btree (d2key)
benchw=# \d dim0
Table "public.dim0"
Column | Type | Modifiers
--------+---------+-----------
d0key | integer | not null
ddate | date | not null
dyr | integer | not null
dmth | integer | not null
dday | integer | not null
Indexes:
"dim0_d0key" UNIQUE, btree (d0key)
The example on the web site has the following execution plan:-
QUERY PLAN
------------------------------------------------------------------------
--------------------
HashAggregate (cost=286953.94..286953.94 rows=1 width=8)
-> Nested Loop (cost=0.00..285268.93 rows=337002 width=8)
-> Seq Scan on dim0 d0 (cost=0.00..219.00 rows=337 width=8)
Filter: ((ddate >= '2010-01-01'::date) AND (ddate <=
'2010-12-28'::date))
-> Index Scan using fact0_d0key on fact0 f (cost=0.00..833.07
rows=1022 width=8)
Index Cond: ("outer".d0key = f.d0key)
It uses the index on the join condition.
When I disable the sequential scan with:-
SET enable_seqscan TO off;
The execution plan looks like:-
QUERY
PLAN
------------------------------------------------------------------------
----------------------------------------------------------------
HashAggregate (cost=648831.52..648831.53 rows=1 width=8) (actual
time=19155.060..19155.071 rows=12 loops=1)
-> Nested Loop (cost=7.51..647176.52 rows=331000 width=8) (actual
time=97.878..18943.155 rows=336000 loops=1)
-> Index Scan using dim0_d0key on dim0 d0 (cost=0.00..248.00
rows=331 width=8) (actual time=40.467..55.780 rows=336 loops=1)
Filter: ((ddate >= '2010-01-01'::date) AND (ddate <=
'2010-12-28'::date))
-> Bitmap Heap Scan on fact0 f (cost=7.51..1941.94 rows=1002
width=8) (actual time=0.991..55.391 rows=1000 loops=336)
Recheck Cond: ("outer".d0key = f.d0key)
-> Bitmap Index Scan on fact0_d0key (cost=0.00..7.51
rows=1002 width=0) (actual time=0.583..0.583 rows=1000 loops=336)
Index Cond: ("outer".d0key = f.d0key)
Total runtime: 19155.176 ms
(9 rows)
The query is 19 seconds long now; down from 34 seconds although the
execution plan doesn't match the example from the website.
Regards
Robin
-----Original Message-----
From: Peter Eisentraut [mailto:peter_e@gmx.net]
Sent: 21 July 2006 12:46
To: pgsql-performance@postgresql.org
Cc: Smith,R,Robin,XJE4JA C
Subject: Re: [PERFORM] Forcing using index instead of sequential scan?
robin.c.smith@bt.com wrote:
> What is the best way to force the use of indexes in these queries?
Well, the brute-force method is to use SET enable_seqscan TO off, but if
you want to get to the bottom of this, you should look at or post the
EXPLAIN ANALYZE output of the offending queries.
--
Peter Eisentraut
http://developer.postgresql.org/~petere/
The tables have all been analysed. I set the work_mem to 500000 and it still doesn't use the index :-( Regards Robin -----Original Message----- From: Richard Huxton [mailto:dev@archonet.com] Sent: 21 July 2006 12:54 To: Smith,R,Robin,XJE4JA C Subject: Re: [PERFORM] Forcing using index instead of sequential scan? robin.c.smith@bt.com wrote: > I have been testing the performance of PostgreSQL using the simple > tool found at http://benchw.sourceforge.net however I have found that > all the queries it run execute with sequential scans. The website > where the code runs has examples of the execution plan using indexes. > > When I disable the sequential plan query 0 and query 1 run faster ( > http://benchw.sourceforge.net/benchw_results_postgres_history.html ) > by using the indexes as suggested by the website. > > I have tried increasing the effective_cache_size and reducing the > random_page_cost to try and force the optimiser to use the index but > it always uses the sequential scan. > > What is the best way to force the use of indexes in these queries? > Currently testing with version 8.1.4. Well, you don't want to be forcing it if possible. Ideally, PG should be able to figure out what to use itself. In the case of query0 and query1 as shown on your web-page I'd expect a sequential scan of dim0 then access via the index on fact0. Reasons why this might not be happening include: 1. Inaccurate stats - ANALYSE your tables 2. Insufficient memory for sorting etc - issue SET work_mem=XXX before the query and try increased values. 3. Other parameters are out-of-whack. For example, effective_cache_size doesn't change how much cache PG uses, it tells PG how much the O.S. will cache. You might find http://www.powerpostgresql.com/PerfList is a good quick introduction. So - ANALYSE your tables http://www.postgresql.org/docs/8.1/static/sql-analyze.html Then post EXPLAIN ANALYSE for the queries and we'll see what they're doing. -- Richard Huxton Archonet Ltd
robin.c.smith@bt.com wrote: > What is the best way to force the use of indexes in these queries? Well, the brute-force method is to use SET enable_seqscan TO off, but if you want to get to the bottom of this, you should look at or post the EXPLAIN ANALYZE output of the offending queries. -- Peter Eisentraut http://developer.postgresql.org/~petere/
<robin.c.smith@bt.com> writes: > I have been testing the performance of PostgreSQL using the simple tool > found at http://benchw.sourceforge.net however I have found that all the > queries it run execute with sequential scans. The website where the code > runs has examples of the execution plan using indexes. The reason the website gets indexscans is that he's fooled with the planner cost parameters. In particular I see that benchw's documentation suggests effective_cache_size = 48000 random_page_cost = 0.8 The latter is physically silly but it's a pretty effective thumb on the scales if you want to force indexscan usage. The real issue here is caching across successive queries, an effect that Postgres doesn't deal with very well at the moment. If you run these queries from a standing start (freshly booted machine) you'll likely find that the indexscan plan is indeed slower than the seqscan/hash plan, just like the planner thinks. I get about 52 sec for query0 with an indexscan vs about 35 sec for the seqscan. However, successive executions of the seqscan plan stay at about 35 sec, whereas the indexscan plan drops to 2 sec(!). This is because the fraction of the table touched by the indexscan plan is small enough to fit in my machine's RAM --- I can see by das blinkenlights (and also vmstat) that there's no I/O going on at all during re-executions of the indexscan. If I run the seqscan and then the indexscan, the indexscan takes about 28 sec, so there's still some useful cached data even though the seqscan read more stuff than fits in RAM. (Note: this is with Fedora Core 5, YMMV depending on your kernel's cache algorithms.) In a real-world situation it's unlikely you'd just re-execute the same query over and over, so this benchmark is really too simplistic to trust very far as an indicator of what to do in practice. I find that CVS tip will choose the indexscan for query0 if I set effective_cache_size to 62500 (ie, half a gigabyte, or half of this machine's RAM) and set random_page_cost to 1.5 or less. If you want the planner to work on the assumption that everything's cached, set effective_cache_size to a large value and set random_page_cost to 1.0 --- you might also want to increase the CPU cost settings, reflecting the fact that I/O is cheaper relative to CPU effort than the default settings assume. However, if your database is too large to fit in RAM then these are likely to be pretty bad settings. Many people compromise with a random_page_cost around 2 or so. regards, tom lane
> The real issue here is caching across successive queries, an effect that
> Postgres doesn't deal with very well at the moment. If you run these
> queries from a standing start (freshly booted machine) you'll likely
> find that the indexscan plan is indeed slower than the seqscan/hash
> plan, just like the planner thinks.
Here's a little trick I learned to speed up this test.
find / -type f -exec grep foobar {} \;
This causes massive file-system activity and flushes all files that the kernel has cached. If you run this between
eachPostgres test (let it run for a couple minutes), it gives you an apples-to-apples comparison between successive
benchmarks,and eliminates the effects of caching.
If you run this as a regular user (NOT super-user or 'postgres'), you won't have permission to access your Postgres
files,so you're guaranteed they'll be flushed from the cache.
Craig
"Craig A. James" <cjames@modgraph-usa.com> writes: > This causes massive file-system activity and flushes all files that the > kernel has cached. If you run this between each Postgres test (let it run > for a couple minutes), it gives you an apples-to-apples comparison between > successive benchmarks, and eliminates the effects of caching. On Linux at least the best way to flush the cache is to unmount and then mount the filesystem. This requires putting the data files on partition that you aren't otherwise using and shutting down postgres. Note that "nothing cached" isn't necessarily any more accurate a model as "everything cached". In reality many databases *do* in fact run the same queries over and over again, though often with some parameters different each time. But the upper pages of most indexes and many of the common leaf pages and heap pages will in fact be cached. -- Gregory Stark EnterpriseDB http://www.enterprisedb.com
Tom Lane wrote: > <robin.c.smith@bt.com> writes: >> I have been testing the performance of PostgreSQL using the simple tool >> found at http://benchw.sourceforge.net however I have found that all the >> queries it run execute with sequential scans. The website where the code >> runs has examples of the execution plan using indexes. > > The reason the website gets indexscans is that he's fooled with the > planner cost parameters. In particular I see that...(snipped) > Indeed I did - probably should have discussed that alteration better in the documentation for the test suite! In addition I was a bit naughty in running the benchmark using size 1 (i.e about 1G) an a box with 2G ram - as this meant that (on the machine I was using then anyway) indexscans on query 0 and 1 were *always* better than the sequential options. A better test is to use the size factor at 2 x physical ram, as then the planners defaults make more sense! (unless or course you *want* to model a data mart smaller than physical ram). Best wishes Mark
robin.c.smith@bt.com wrote: > I have been testing the performance of PostgreSQL using the simple tool > found at _http://benchw.sourceforge.net_ however I have found that all > the queries it run execute with sequential scans. The website where the > code runs has examples of the execution plan using indexes. > > When I disable the sequential plan query 0 and query 1 run faster ( > _http://benchw.sourceforge.net/benchw_results_postgres_history.html_ ) > by using the indexes as suggested by the website. > > I have tried increasing the effective_cache_size and reducing the > random_page_cost to try and force the optimiser to use the index but it > always uses the sequential scan. > > What is the best way to force the use of indexes in these queries? > Currently testing with version 8.1.4. > > Hi Robin, Being responsible for this piece of software, I should try to help, only saw this now sorry (nice to see someone using this). Unless you really want to reproduce the numbers on the website, it is best to test with Benchw's scale factor at least 2 x your physical ram, as this makes the planner's defaults work more sensibly (and models *most* real world data warehouse situations better!). Cheers Mark
On Sat, Jul 22, 2006 at 10:26:53AM -0700, Craig A. James wrote: >This causes massive file-system activity and flushes all files that the >kernel has cached. If you run this between each Postgres test (let it run >for a couple minutes), it gives you an apples-to-apples comparison between >successive benchmarks, and eliminates the effects of caching. Assuming a system with small ram or an unusually large system installation. Unmounting is a much more realiable mechanism. Mike Stone
Michael Stone wrote: > On Sat, Jul 22, 2006 at 10:26:53AM -0700, Craig A. James wrote: >> This causes massive file-system activity and flushes all files that >> the kernel has cached. If you run this between each Postgres test >> (let it run for a couple minutes), it gives you an apples-to-apples >> comparison between successive benchmarks, and eliminates the effects >> of caching. > > Assuming a system with small ram or an unusually large system > installation. Unmounting is a much more realiable mechanism. Indeed, but it only works if you can. For example, in my small-ish installation, my WAL and system tables are mounted onthe root disk. Or someone might not have super-user access. Craig