Обсуждение: Creating table with data from a join
Hello,
default_statistics_target = 50 # pgtune wizard 2012-06-06 maintenance_work_mem = 1GB # pgtune wizard 2012-06-06 constraint_exclusion = on # pgtune wizard 2012-06-06 checkpoint_completion_target = 0.9 # pgtune wizard 2012-06-06 effective_cache_size = 48GB # pgtune wizard 2012-06-06
I am benchmarking different ways of putting data into table on table creation:
parallel --block 128M --jobs 4 --pipe psql -c "COPY c FROM STDIN";
(the parallel command is available as part of parallel deb package in Ubuntu for example, it splits the stdin by newline character and feeds it to the corresponding command)
Both tables a and b have ~16M records and one of the columns in a is geometry (ranging from several KB in size to several MB). Columns in b are mostly integers.
The machine that I am running these commands on has the following parameters:
work_mem = 80MB # pgtune wizard 2012-06-06
wal_buffers = 8MB # pgtune wizard 2012-06-06
checkpoint_segments = 16 # pgtune wizard 2012-06-06
shared_buffers = 16GB # pgtune wizard 2012-06-06
max_connections = 400 # pgtune wizard 2012-06-06
One would expect the 3rd option to be faster than 1 and 2, however 2 outperforms both by a large margin (sometimes x2). This is especially surprising taking into account that COPY doesn't acquire a global lock on the table, only a RowExclusiveLock
So is option 2 a winner by design? Could you please suggest other alternatives to try (if there are any)? And what might be the reason that 3 is not outperforming the other 2?
Thank you,
Igor
On 14 July 2015 at 21:12, Igor Stassiy <istassiy@gmail.com> wrote:
Hello,I am benchmarking different ways of putting data into table on table creation:parallel --block 128M --jobs 4 --pipe psql -c "COPY c FROM STDIN";(the parallel command is available as part of parallel deb package in Ubuntu for example, it splits the stdin by newline character and feeds it to the corresponding command)Both tables a and b have ~16M records and one of the columns in a is geometry (ranging from several KB in size to several MB). Columns in b are mostly integers.The machine that I am running these commands on has the following parameters:default_statistics_target = 50 # pgtune wizard 2012-06-06 maintenance_work_mem = 1GB # pgtune wizard 2012-06-06 constraint_exclusion = on # pgtune wizard 2012-06-06 checkpoint_completion_target = 0.9 # pgtune wizard 2012-06-06 effective_cache_size = 48GB # pgtune wizard 2012-06-06work_mem = 80MB # pgtune wizard 2012-06-06wal_buffers = 8MB # pgtune wizard 2012-06-06checkpoint_segments = 16 # pgtune wizard 2012-06-06shared_buffers = 16GB # pgtune wizard 2012-06-06max_connections = 400 # pgtune wizard 2012-06-06One would expect the 3rd option to be faster than 1 and 2, however 2 outperforms both by a large margin (sometimes x2). This is especially surprising taking into account that COPY doesn't acquire a global lock on the table, only a RowExclusiveLockSo is option 2 a winner by design? Could you please suggest other alternatives to try (if there are any)? And what might be the reason that 3 is not outperforming the other 2?
I would imagine that the calling of the output function to translate the each value's internal representation to it's user visible/external representation plus all the overhead of sending results to the client would be a likely candidate of the slow down. In either case 3 would only be as fast as the query generating the output. With 1 and 2 all the tuple representations of each record stays in the internal format.
If you have some logical way to break the query down into parts, then maybe that would be a place to look.
For example:
Of course, you'd need to be very careful to ensure that the results of each SELECT never overlap. It would be nice to invent some better way than this that divided the workload evenly even when the tables grow.
Then you could run these concurrently.
Regards
David Rowley
--
David Rowley http://www.2ndQuadrant.com/
David Rowley http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services
Thank you David. I tried to measure the time of COPY (SELECT * FROM a JOIN b ON a.id = b.id) TO '/tmp/dump.sql' and it took an order of magnitude time less (~10x) than the complete command (together with INSERT), so conversion is probably not the main factor of slowdown (unless conversion from text ->internal is significantly slower than that of from internal -> text).
I will also try your suggestion with limiting the ids range.
On Tue, Jul 14, 2015 at 1:42 PM David Rowley <david.rowley@2ndquadrant.com> wrote:
On 14 July 2015 at 21:12, Igor Stassiy <istassiy@gmail.com> wrote:Hello,I am benchmarking different ways of putting data into table on table creation:parallel --block 128M --jobs 4 --pipe psql -c "COPY c FROM STDIN";(the parallel command is available as part of parallel deb package in Ubuntu for example, it splits the stdin by newline character and feeds it to the corresponding command)Both tables a and b have ~16M records and one of the columns in a is geometry (ranging from several KB in size to several MB). Columns in b are mostly integers.The machine that I am running these commands on has the following parameters:default_statistics_target = 50 # pgtune wizard 2012-06-06 maintenance_work_mem = 1GB # pgtune wizard 2012-06-06 constraint_exclusion = on # pgtune wizard 2012-06-06 checkpoint_completion_target = 0.9 # pgtune wizard 2012-06-06 effective_cache_size = 48GB # pgtune wizard 2012-06-06work_mem = 80MB # pgtune wizard 2012-06-06wal_buffers = 8MB # pgtune wizard 2012-06-06checkpoint_segments = 16 # pgtune wizard 2012-06-06shared_buffers = 16GB # pgtune wizard 2012-06-06max_connections = 400 # pgtune wizard 2012-06-06One would expect the 3rd option to be faster than 1 and 2, however 2 outperforms both by a large margin (sometimes x2). This is especially surprising taking into account that COPY doesn't acquire a global lock on the table, only a RowExclusiveLockSo is option 2 a winner by design? Could you please suggest other alternatives to try (if there are any)? And what might be the reason that 3 is not outperforming the other 2?I would imagine that the calling of the output function to translate the each value's internal representation to it's user visible/external representation plus all the overhead of sending results to the client would be a likely candidate of the slow down. In either case 3 would only be as fast as the query generating the output. With 1 and 2 all the tuple representations of each record stays in the internal format.If you have some logical way to break the query down into parts, then maybe that would be a place to look.For example:Of course, you'd need to be very careful to ensure that the results of each SELECT never overlap. It would be nice to invent some better way than this that divided the workload evenly even when the tables grow.Then you could run these concurrently.RegardsDavid Rowley--
David Rowley http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training & Services
On 14/07/2015 11:12, Igor Stassiy wrote: > Hello, > > I am benchmarking different ways of putting data into table on table > creation: > > 1. INSERT INTO c SELECT * FROM a JOIN b on a.id <http://a.id> = b.id > <http://b.id>; > 2. CREATE TABLE c AS SELECT * FROM a JOIN b on a.id <http://a.id> = b.id > <http://b.id>; > 3. psql -c "COPY (SELECT * FROM a JOIN b on a.id <http://a.id> = b.id > <http://b.id>) TO STDOUT" | > parallel --block 128M --jobs 4 --pipe psql -c "COPY c FROM STDIN"; > > (the parallel command is available as part of parallel deb package in > Ubuntu for example, it splits the stdin by newline character and feeds > it to the corresponding command) > > Both tables a and b have ~16M records and one of the columns in a is > geometry (ranging from several KB in size to several MB). Columns in b > are mostly integers. > > The machine that I am running these commands on has the following > parameters: > > default_statistics_target = 50 # pgtune wizard 2012-06-06 > maintenance_work_mem = 1GB # pgtune wizard 2012-06-06 > constraint_exclusion = on # pgtune wizard 2012-06-06 > checkpoint_completion_target = 0.9 # pgtune wizard 2012-06-06 > effective_cache_size = 48GB # pgtune wizard 2012-06-06 > work_mem = 80MB # pgtune wizard 2012-06-06 > wal_buffers = 8MB # pgtune wizard 2012-06-06 > checkpoint_segments = 16 # pgtune wizard 2012-06-06 > shared_buffers = 16GB # pgtune wizard 2012-06-06 > max_connections = 400 # pgtune wizard 2012-06-06 > > One would expect the 3rd option to be faster than 1 and 2, however 2 > outperforms both by a large margin (sometimes x2). This is especially > surprising taking into account that COPY doesn't acquire a global lock > on the table, only a RowExclusiveLock > (according > to http://www.postgresql.org/message-id/10611.1014867684@sss.pgh.pa.us) > What is wal_level value? I think this is because of an optimisation happening with wal_level = minimal: "In minimal level, WAL-logging of some bulk operations can be safely skipped, which can make those operations much faster" see http://www.postgresql.org/docs/current/static/runtime-config-wal.html > So is option 2 a winner by design? Could you please suggest other > alternatives to try (if there are any)? And what might be the reason > that 3 is not outperforming the other 2? > > Thank you, > Igor > > -- Julien Rouhaud http://dalibo.com - http://dalibo.org
Julien, I have the following setting for WAL level: #wal_level = minimal (which defaults to minimal anyway)
On Tue, Jul 14, 2015 at 6:19 PM Julien Rouhaud <julien.rouhaud@dalibo.com> wrote:
On 14/07/2015 11:12, Igor Stassiy wrote:
> Hello,
>
> I am benchmarking different ways of putting data into table on table
> creation:
>
> 1. INSERT INTO c SELECT * FROM a JOIN b on a.id <http://a.id> = b.id
> <http://b.id>;
> 2. CREATE TABLE c AS SELECT * FROM a JOIN b on a.id <http://a.id> = b.id
> <http://b.id>;
> 3. psql -c "COPY (SELECT * FROM a JOIN b on a.id <http://a.id> = b.id
> <http://b.id>) TO STDOUT" |
> parallel --block 128M --jobs 4 --pipe psql -c "COPY c FROM STDIN";
>
> (the parallel command is available as part of parallel deb package in
> Ubuntu for example, it splits the stdin by newline character and feeds
> it to the corresponding command)
>
> Both tables a and b have ~16M records and one of the columns in a is
> geometry (ranging from several KB in size to several MB). Columns in b
> are mostly integers.
>
> The machine that I am running these commands on has the following
> parameters:
>
> default_statistics_target = 50 # pgtune wizard 2012-06-06
> maintenance_work_mem = 1GB # pgtune wizard 2012-06-06
> constraint_exclusion = on # pgtune wizard 2012-06-06
> checkpoint_completion_target = 0.9 # pgtune wizard 2012-06-06
> effective_cache_size = 48GB # pgtune wizard 2012-06-06
> work_mem = 80MB # pgtune wizard 2012-06-06
> wal_buffers = 8MB # pgtune wizard 2012-06-06
> checkpoint_segments = 16 # pgtune wizard 2012-06-06
> shared_buffers = 16GB # pgtune wizard 2012-06-06
> max_connections = 400 # pgtune wizard 2012-06-06
>
> One would expect the 3rd option to be faster than 1 and 2, however 2
> outperforms both by a large margin (sometimes x2). This is especially
> surprising taking into account that COPY doesn't acquire a global lock
> on the table, only a RowExclusiveLock
> (according
> to http://www.postgresql.org/message-id/10611.1014867684@sss.pgh.pa.us)
>
What is wal_level value? I think this is because of an optimisation
happening with wal_level = minimal:
"In minimal level, WAL-logging of some bulk operations can be safely
skipped, which can make those operations much faster"
see http://www.postgresql.org/docs/current/static/runtime-config-wal.html
> So is option 2 a winner by design? Could you please suggest other
> alternatives to try (if there are any)? And what might be the reason
> that 3 is not outperforming the other 2?
>
> Thank you,
> Igor
>
>
--
Julien Rouhaud
http://dalibo.com - http://dalibo.org
-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1 On 14/07/2015 18:21, Igor Stassiy wrote: > Julien, I have the following setting for WAL level: #wal_level = > minimal (which defaults to minimal anyway) > > On Tue, Jul 14, 2015 at 6:19 PM Julien Rouhaud > <julien.rouhaud@dalibo.com <mailto:julien.rouhaud@dalibo.com>> > wrote: > > On 14/07/2015 11:12, Igor Stassiy wrote: >> Hello, >> >> I am benchmarking different ways of putting data into table on >> table creation: >> >> 1. INSERT INTO c SELECT * FROM a JOIN b on a.id <http://a.id> > <http://a.id> = b.id <http://b.id> >> <http://b.id>; 2. CREATE TABLE c AS SELECT * FROM a JOIN b on >> a.id <http://a.id> > <http://a.id> = b.id <http://b.id> >> <http://b.id>; 3. psql -c "COPY (SELECT * FROM a JOIN b on a.id >> <http://a.id> > <http://a.id> = b.id <http://b.id> >> <http://b.id>) TO STDOUT" | parallel --block 128M --jobs 4 --pipe >> psql -c "COPY c FROM STDIN"; >> >> (the parallel command is available as part of parallel deb >> package in Ubuntu for example, it splits the stdin by newline >> character and feeds it to the corresponding command) >> >> Both tables a and b have ~16M records and one of the columns in a >> is geometry (ranging from several KB in size to several MB). >> Columns in b are mostly integers. >> >> The machine that I am running these commands on has the >> following parameters: >> >> default_statistics_target = 50 # pgtune wizard 2012-06-06 >> maintenance_work_mem = 1GB # pgtune wizard 2012-06-06 >> constraint_exclusion = on # pgtune wizard 2012-06-06 >> checkpoint_completion_target = 0.9 # pgtune wizard 2012-06-06 >> effective_cache_size = 48GB # pgtune wizard 2012-06-06 work_mem = >> 80MB # pgtune wizard 2012-06-06 wal_buffers = 8MB # pgtune wizard >> 2012-06-06 checkpoint_segments = 16 # pgtune wizard 2012-06-06 >> shared_buffers = 16GB # pgtune wizard 2012-06-06 max_connections >> = 400 # pgtune wizard 2012-06-06 >> >> One would expect the 3rd option to be faster than 1 and 2, >> however 2 outperforms both by a large margin (sometimes x2). This >> is especially surprising taking into account that COPY doesn't >> acquire a global lock on the table, only a RowExclusiveLock >> (according to > http://www.postgresql.org/message-id/10611.1014867684@sss.pgh.pa.us) > > > > > What is wal_level value? I think this is because of an > optimisation happening with wal_level = minimal: > > "In minimal level, WAL-logging of some bulk operations can be > safely skipped, which can make those operations much faster" > > see > http://www.postgresql.org/docs/current/static/runtime-config-wal.html > > >> So is option 2 a winner by design? Could you please suggest >> other alternatives to try (if there are any)? And what might be >> the reason that 3 is not outperforming the other 2? >> >> Thank you, Igor >> >> > > > -- Julien Rouhaud http://dalibo.com - http://dalibo.org > - -- Julien Rouhaud http://dalibo.com - http://dalibo.org -----BEGIN PGP SIGNATURE----- Version: GnuPG v2.0.17 (GNU/Linux) iQEcBAEBAgAGBQJVpTlRAAoJELGaJ8vfEpOqvI4H/RZygc5QXOuEZDWqmWRoZZ5N kNLWxPJbQ7cLpSNIUj3gJmq9bj0I3K071L09KbJWgxtwvQCzgiTsUIVURv7V83C6 nQ8CmrRr96+jKprx5Gw/uqSel8qnbi9LApl1IDqx9Hnd/HnyVOemND2gzHOQhsKN tvGuo4ac5yR+rsFA8FHuwXgSgVH2NEDL2n4Zv6jI2uwh5NRBeeGEn8MFKDZCSWN6 HXG9wZaelSrYbcSfumRg07RLnAmP6E/xbY1eB8dA17XmnFxE9AMTFy0YqJb8Kl5Z KvzQ6+VHnrW2zaoCUOGE56ra2La7TPeJxxeNA9U9Li+8GmvJIQHqIoQvLz7CzT8= =Ztkl -----END PGP SIGNATURE-----
On 14/07/2015 18:21, Igor Stassiy wrote: > Julien, I have the following setting for WAL level: #wal_level = minimal > (which defaults to minimal anyway) > Sorry, I sent my mail too early :/ So, option #2 is winner by design. You didn't say anything about your needs, so it's hard to help you much more. If you don't care about losing data on this table if your server crashes, you can try option #3 with an unlogged table. > On Tue, Jul 14, 2015 at 6:19 PM Julien Rouhaud > <julien.rouhaud@dalibo.com <mailto:julien.rouhaud@dalibo.com>> wrote: > > On 14/07/2015 11:12, Igor Stassiy wrote: > > Hello, > > > > I am benchmarking different ways of putting data into table on table > > creation: > > > > 1. INSERT INTO c SELECT * FROM a JOIN b on a.id <http://a.id> > <http://a.id> = b.id <http://b.id> > > <http://b.id>; > > 2. CREATE TABLE c AS SELECT * FROM a JOIN b on a.id <http://a.id> > <http://a.id> = b.id <http://b.id> > > <http://b.id>; > > 3. psql -c "COPY (SELECT * FROM a JOIN b on a.id <http://a.id> > <http://a.id> = b.id <http://b.id> > > <http://b.id>) TO STDOUT" | > > parallel --block 128M --jobs 4 --pipe psql -c "COPY c FROM STDIN"; > > > > (the parallel command is available as part of parallel deb package in > > Ubuntu for example, it splits the stdin by newline character and feeds > > it to the corresponding command) > > > > Both tables a and b have ~16M records and one of the columns in a is > > geometry (ranging from several KB in size to several MB). Columns in b > > are mostly integers. > > > > The machine that I am running these commands on has the following > > parameters: > > > > default_statistics_target = 50 # pgtune wizard 2012-06-06 > > maintenance_work_mem = 1GB # pgtune wizard 2012-06-06 > > constraint_exclusion = on # pgtune wizard 2012-06-06 > > checkpoint_completion_target = 0.9 # pgtune wizard 2012-06-06 > > effective_cache_size = 48GB # pgtune wizard 2012-06-06 > > work_mem = 80MB # pgtune wizard 2012-06-06 > > wal_buffers = 8MB # pgtune wizard 2012-06-06 > > checkpoint_segments = 16 # pgtune wizard 2012-06-06 > > shared_buffers = 16GB # pgtune wizard 2012-06-06 > > max_connections = 400 # pgtune wizard 2012-06-06 > > > > One would expect the 3rd option to be faster than 1 and 2, however 2 > > outperforms both by a large margin (sometimes x2). This is especially > > surprising taking into account that COPY doesn't acquire a global lock > > on the table, only a RowExclusiveLock > > (according > > to > http://www.postgresql.org/message-id/10611.1014867684@sss.pgh.pa.us) > > > > What is wal_level value? I think this is because of an optimisation > happening with wal_level = minimal: > > "In minimal level, WAL-logging of some bulk operations can be safely > skipped, which can make those operations much faster" > > see > http://www.postgresql.org/docs/current/static/runtime-config-wal.html > > > So is option 2 a winner by design? Could you please suggest other > > alternatives to try (if there are any)? And what might be the reason > > that 3 is not outperforming the other 2? > > > > Thank you, > > Igor > > > > > > > -- > Julien Rouhaud > http://dalibo.com - http://dalibo.org > -- Julien Rouhaud http://dalibo.com - http://dalibo.org
Julien, I would gladly provide more information, I am just not sure what to add.
On Tue, Jul 14, 2015 at 6:37 PM Julien Rouhaud <julien.rouhaud@dalibo.com> wrote:
On 14/07/2015 18:21, Igor Stassiy wrote:
> Julien, I have the following setting for WAL level: #wal_level = minimal
> (which defaults to minimal anyway)
>
Sorry, I sent my mail too early :/
So, option #2 is winner by design. You didn't say anything about your
needs, so it's hard to help you much more.
If you don't care about losing data on this table if your server
crashes, you can try option #3 with an unlogged table.
> On Tue, Jul 14, 2015 at 6:19 PM Julien Rouhaud
> <julien.rouhaud@dalibo.com <mailto:julien.rouhaud@dalibo.com>> wrote:
>
> On 14/07/2015 11:12, Igor Stassiy wrote:
> > Hello,
> >
> > I am benchmarking different ways of putting data into table on table
> > creation:
> >
> > 1. INSERT INTO c SELECT * FROM a JOIN b on a.id <http://a.id>
> <http://a.id> = b.id <http://b.id>
> > <http://b.id>;
> > 2. CREATE TABLE c AS SELECT * FROM a JOIN b on a.id <http://a.id>
> <http://a.id> = b.id <http://b.id>
> > <http://b.id>;
> > 3. psql -c "COPY (SELECT * FROM a JOIN b on a.id <http://a.id>
> <http://a.id> = b.id <http://b.id>
> > <http://b.id>) TO STDOUT" |
> > parallel --block 128M --jobs 4 --pipe psql -c "COPY c FROM STDIN";
> >
> > (the parallel command is available as part of parallel deb package in
> > Ubuntu for example, it splits the stdin by newline character and feeds
> > it to the corresponding command)
> >
> > Both tables a and b have ~16M records and one of the columns in a is
> > geometry (ranging from several KB in size to several MB). Columns in b
> > are mostly integers.
> >
> > The machine that I am running these commands on has the following
> > parameters:
> >
> > default_statistics_target = 50 # pgtune wizard 2012-06-06
> > maintenance_work_mem = 1GB # pgtune wizard 2012-06-06
> > constraint_exclusion = on # pgtune wizard 2012-06-06
> > checkpoint_completion_target = 0.9 # pgtune wizard 2012-06-06
> > effective_cache_size = 48GB # pgtune wizard 2012-06-06
> > work_mem = 80MB # pgtune wizard 2012-06-06
> > wal_buffers = 8MB # pgtune wizard 2012-06-06
> > checkpoint_segments = 16 # pgtune wizard 2012-06-06
> > shared_buffers = 16GB # pgtune wizard 2012-06-06
> > max_connections = 400 # pgtune wizard 2012-06-06
> >
> > One would expect the 3rd option to be faster than 1 and 2, however 2
> > outperforms both by a large margin (sometimes x2). This is especially
> > surprising taking into account that COPY doesn't acquire a global lock
> > on the table, only a RowExclusiveLock
> > (according
> > to
> http://www.postgresql.org/message-id/10611.1014867684@sss.pgh.pa.us)
> >
>
> What is wal_level value? I think this is because of an optimisation
> happening with wal_level = minimal:
>
> "In minimal level, WAL-logging of some bulk operations can be safely
> skipped, which can make those operations much faster"
>
> see
> http://www.postgresql.org/docs/current/static/runtime-config-wal.html
>
> > So is option 2 a winner by design? Could you please suggest other
> > alternatives to try (if there are any)? And what might be the reason
> > that 3 is not outperforming the other 2?
> >
> > Thank you,
> > Igor
> >
> >
>
>
> --
> Julien Rouhaud
> http://dalibo.com - http://dalibo.org
>
--
Julien Rouhaud
http://dalibo.com - http://dalibo.org
>Hello,
>
>I am benchmarking different ways of putting data into table on table creation:
>
>1. INSERT INTO c SELECT * FROM a JOIN b on a.id = b.id;
>2. CREATE TABLE c AS SELECT * FROM a JOIN b on a.id = b.id;
>3. psql -c "COPY (SELECT * FROM a JOIN b on a.id = b.id) TO STDOUT" |
>parallel --block 128M --jobs 4 --pipe psql -c "COPY c FROM STDIN";
>
>(the parallel command is available as part of parallel deb package in Ubuntu for example, it splits the stdin by newline character and feeds it to the corresponding command)
>
>Both tables a and b have ~16M records and one of the columns in a is geometry (ranging from several KB in size to several MB). Columns in b are mostly integers.
>
>The machine that I am running these commands on has the following parameters:
>
>default_statistics_target = 50 # pgtune wizard 2012-06-06 maintenance_work_mem = 1GB # pgtune wizard 2012-06-06 constraint_exclusion = on # pgtune wizard 2012-06-06 checkpoint_completion_target = 0.9 # pgtune wizard 2012-06-06 effective_cache_size = 48GB # pgtune wizard 2012-06-06
>work_mem = 80MB # pgtune wizard 2012-06-06
>wal_buffers = 8MB # pgtune wizard 2012-06-06
>checkpoint_segments = 16 # pgtune wizard 2012-06-06
>shared_buffers = 16GB # pgtune wizard 2012-06-06
>max_connections = 400 # pgtune wizard 2012-06-06
>
>One would expect the 3rd option to be faster than 1 and 2, however 2 outperforms both by a large margin (sometimes x2). This is especially surprising taking into account that COPY doesn't acquire a global lock on the table, only a RowExclusiveLock
which PG Version ?
I find interesting, that 2 outperforms 1.
The only explanation I can imagine is that "CREATE TABLE AS" freezes the data on the fly, as possible with "COPY FROM"
(http://www.postgresql.org/docs/9.4/interactive/sql-copy.html)
You may try parallel insert without using STDIN using modulo. Just start these 4 queries simultaneously:
INSERT INTO c SELECT * FROM a JOIN b on a.id = b.id
WHERE a.id%4 = [0,1,2,3}
I usually avoid parallel INSERTS to avoid I/O contention and random distribution within the target tables.
Are you monitoring the I/O activity in your tests ?
Have you tried to use only 2 parallel processes?
regards,
Marc Mamin
>(according to http://www.postgresql.org/message-id/10611.1014867684@sss.pgh.pa.us)
>
>So is option 2 a winner by design? Could you please suggest other alternatives to try (if there are any)? And what might be the reason that 3 is not outperforming the other 2?
>
>Thank you,
>Igor
>
>
>
On 14/07/2015 18:50, Igor Stassiy wrote: > Julien, I would gladly provide more information, I am just not sure what > to add. > Well, was your concern about why option #2 is the quickest, or is this runtime with option #2 still too slow for you ? > I would be willing to leave the server compromised for things like > corrupts or data losses during the time of this import, but the server > has to be up and running before and after the import, if it is > successful (so I can't take it down then change some parameters and > start it up with again). Check http://www.postgresql.org/docs/current/static/sql-createtable.html and the "UNLOGGED" part to check if an unlogged table is suitable for you. > > On Tue, Jul 14, 2015 at 6:37 PM Julien Rouhaud > <julien.rouhaud@dalibo.com <mailto:julien.rouhaud@dalibo.com>> wrote: > > On 14/07/2015 18:21, Igor Stassiy wrote: > > Julien, I have the following setting for WAL level: #wal_level = > minimal > > (which defaults to minimal anyway) > > > > Sorry, I sent my mail too early :/ > > So, option #2 is winner by design. You didn't say anything about your > needs, so it's hard to help you much more. > > If you don't care about losing data on this table if your server > crashes, you can try option #3 with an unlogged table. > > > > On Tue, Jul 14, 2015 at 6:19 PM Julien Rouhaud > > <julien.rouhaud@dalibo.com <mailto:julien.rouhaud@dalibo.com> > <mailto:julien.rouhaud@dalibo.com > <mailto:julien.rouhaud@dalibo.com>>> wrote: > > > > On 14/07/2015 11:12, Igor Stassiy wrote: > > > Hello, > > > > > > I am benchmarking different ways of putting data into table > on table > > > creation: > > > > > > 1. INSERT INTO c SELECT * FROM a JOIN b on a.id > <http://a.id> <http://a.id> > > <http://a.id> = b.id <http://b.id> <http://b.id> > > > <http://b.id>; > > > 2. CREATE TABLE c AS SELECT * FROM a JOIN b on a.id > <http://a.id> <http://a.id> > > <http://a.id> = b.id <http://b.id> <http://b.id> > > > <http://b.id>; > > > 3. psql -c "COPY (SELECT * FROM a JOIN b on a.id > <http://a.id> <http://a.id> > > <http://a.id> = b.id <http://b.id> <http://b.id> > > > <http://b.id>) TO STDOUT" | > > > parallel --block 128M --jobs 4 --pipe psql -c "COPY c FROM > STDIN"; > > > > > > (the parallel command is available as part of parallel deb > package in > > > Ubuntu for example, it splits the stdin by newline character > and feeds > > > it to the corresponding command) > > > > > > Both tables a and b have ~16M records and one of the columns > in a is > > > geometry (ranging from several KB in size to several MB). > Columns in b > > > are mostly integers. > > > > > > The machine that I am running these commands on has the > following > > > parameters: > > > > > > default_statistics_target = 50 # pgtune wizard 2012-06-06 > > > maintenance_work_mem = 1GB # pgtune wizard 2012-06-06 > > > constraint_exclusion = on # pgtune wizard 2012-06-06 > > > checkpoint_completion_target = 0.9 # pgtune wizard 2012-06-06 > > > effective_cache_size = 48GB # pgtune wizard 2012-06-06 > > > work_mem = 80MB # pgtune wizard 2012-06-06 > > > wal_buffers = 8MB # pgtune wizard 2012-06-06 > > > checkpoint_segments = 16 # pgtune wizard 2012-06-06 > > > shared_buffers = 16GB # pgtune wizard 2012-06-06 > > > max_connections = 400 # pgtune wizard 2012-06-06 > > > > > > One would expect the 3rd option to be faster than 1 and 2, > however 2 > > > outperforms both by a large margin (sometimes x2). This is > especially > > > surprising taking into account that COPY doesn't acquire a > global lock > > > on the table, only a RowExclusiveLock > > > (according > > > to > > > http://www.postgresql.org/message-id/10611.1014867684@sss.pgh.pa.us) > > > > > > > What is wal_level value? I think this is because of an > optimisation > > happening with wal_level = minimal: > > > > "In minimal level, WAL-logging of some bulk operations can be > safely > > skipped, which can make those operations much faster" > > > > see > > > http://www.postgresql.org/docs/current/static/runtime-config-wal.html > > > > > So is option 2 a winner by design? Could you please suggest > other > > > alternatives to try (if there are any)? And what might be > the reason > > > that 3 is not outperforming the other 2? > > > > > > Thank you, > > > Igor > > > > > > > > > > > > -- > > Julien Rouhaud > > http://dalibo.com - http://dalibo.org > > > > > -- > Julien Rouhaud > http://dalibo.com - http://dalibo.org > -- Julien Rouhaud http://dalibo.com - http://dalibo.org
David, I did something like this:
Julien, my concern was why the option 3 (with parallel) is not the fastest. And now, even with parallel INSERT INTO .. SELECT its not the fastest. I can't really use the UNLOGGED table in this case.
psql -f /dev/fd/3 3 << IN1 & psql -f /dev/fd/4 4 << IN2 ...
IN1
INSERT INTO c SELECT * FRO a JOIN b ON a.ad=b.id WHERE a.id < 0.5th quantile AND a.id >= 0.25th quantile
IN2
...
IN3
...
IN4
And quantiles were computed using the function:
And quantiles were computed using the function:
SELECT percentile_cont(0.5) WITHIN GROUP (ORDER BY id) FROM a, and then once more for lower and upper halves. But unfortunately, I got only about 16% improvement from non-parallelized version of INSERT INTO .. SELECT ..
Marc, I am using postgres 9.4. I didn't benchmark, but intuitively the modulo operator will force traversing every record in table "a" 4 times, as it can't use an index.
Julien, my concern was why the option 3 (with parallel) is not the fastest. And now, even with parallel INSERT INTO .. SELECT its not the fastest. I can't really use the UNLOGGED table in this case.
The following document summarises why is CREATE TABLE AS .. the fastest:
Basically CREATE TABLE AS .. just doesn't write to wal if the wal_level is minimal and hence cuts IO about in half.
On Tue, Jul 14, 2015 at 1:42 PM David Rowley <david.rowley@2ndquadrant.com> wrote:
On 14 July 2015 at 21:12, Igor Stassiy <istassiy@gmail.com> wrote:Hello,I am benchmarking different ways of putting data into table on table creation:parallel --block 128M --jobs 4 --pipe psql -c "COPY c FROM STDIN";(the parallel command is available as part of parallel deb package in Ubuntu for example, it splits the stdin by newline character and feeds it to the corresponding command)Both tables a and b have ~16M records and one of the columns in a is geometry (ranging from several KB in size to several MB). Columns in b are mostly integers.The machine that I am running these commands on has the following parameters:default_statistics_target = 50 # pgtune wizard 2012-06-06 maintenance_work_mem = 1GB # pgtune wizard 2012-06-06 constraint_exclusion = on # pgtune wizard 2012-06-06 checkpoint_completion_target = 0.9 # pgtune wizard 2012-06-06 effective_cache_size = 48GB # pgtune wizard 2012-06-06work_mem = 80MB # pgtune wizard 2012-06-06wal_buffers = 8MB # pgtune wizard 2012-06-06checkpoint_segments = 16 # pgtune wizard 2012-06-06shared_buffers = 16GB # pgtune wizard 2012-06-06max_connections = 400 # pgtune wizard 2012-06-06One would expect the 3rd option to be faster than 1 and 2, however 2 outperforms both by a large margin (sometimes x2). This is especially surprising taking into account that COPY doesn't acquire a global lock on the table, only a RowExclusiveLockSo is option 2 a winner by design? Could you please suggest other alternatives to try (if there are any)? And what might be the reason that 3 is not outperforming the other 2?I would imagine that the calling of the output function to translate the each value's internal representation to it's user visible/external representation plus all the overhead of sending results to the client would be a likely candidate of the slow down. In either case 3 would only be as fast as the query generating the output. With 1 and 2 all the tuple representations of each record stays in the internal format.If you have some logical way to break the query down into parts, then maybe that would be a place to look.For example:Of course, you'd need to be very careful to ensure that the results of each SELECT never overlap. It would be nice to invent some better way than this that divided the workload evenly even when the tables grow.Then you could run these concurrently.RegardsDavid Rowley--
David Rowley http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training & Services
> Marc, I am using postgres 9.4. I didn't benchmark, but intuitively the modulo operator will force traversing every recordin table "a" 4 times, as it can't use an index. Not necessarily. seq scans can be synchronized: "This allows sequential scans of large tables to synchronize with each other, so that concurrent scans read the same blockat about the same time and hence share the I/O workload. When this is enabled, a scan might start in the middle of thetable and then "wrap around" the end to cover all rows, so as to synchronize with the activity of scans already in progress.This can result in unpredictable changes in the row ordering returned by queries that have no ORDER BY clause." Marc