Re: Performance problem with large insert statements

In looking at some performance issues (I was trying to look at the
overhead of toast) I found that large insert statements were very slow.

My test case involved reading a file (6M file in my tests) and inserting
it into the database into a "largeobject" like table defined as follows:

create table tblob1 (filename text, lastbyte integer, data bytea);

The first test program read the file 8000 bytes at a time and inserted
them into the above table until the entire file was inserted. This test
program used a regular insert statement to do the inserting: (insert
into tblob1 values (?,?,?))

For three runs of this test the average time to insert the 6M file into
the database in 8000 byte rows (which ended up being 801 rows inserted
into the table) was: 17.803 seconds

The second test read the file in 8000 byte chucks just like the first
program but it used a function to do the insert and called the function
via the FastPath API. The function was:

CREATE FUNCTION BYTEA_WRITE (TEXT, INTEGER, BYTEA) RETURNS INTEGER
as '
BEGIN
INSERT INTO TBLOB1 VALUES ($1, $2, $3);
RETURN 1;
END;'
language 'plpgsql'

For three runs of this test the average time to insert the 6M file into
the database in 8000 byte parts was: 2.645

Thus using the insert statement was almost an order of magnitude slower
that using the function (17.803 sec vs. 2.645 sec).

Reading the data back from the server via a standard select statement
takes on average: 1.674 seconds.

I tried to run gprof to see where the time was going, but for some
reason the gprof output on my gmon.out file doesn't have any timing
information (all times are reported as 0.0) and I haven't been able to
figure out why yet. So I don't know what is taking up the bulk of the
time (I suspect it is either the decoding of the bytea data which the
Fastpath function call avoids, or the parser which needs to parse 801 8K
SQL statements vs. the function which has to parse one 100 byte statement.)

I have attached the two test programs (they are in java and use jdbc)
and a SQL script that creates the table and function.

thanks,
--Barry

Barry Lind <barry(at)xythos(dot)com> writes:
> In looking at some performance issues (I was trying to look at the
> overhead of toast) I found that large insert statements were very slow.
> ...
> I tried to run gprof to see where the time was going, but for some
> reason the gprof output on my gmon.out file doesn't have any timing
> information (all times are reported as 0.0) and I haven't been able to
> figure out why yet.

That seems to be a common disease among Linuxen; dunno why. gprof
works fine for me on HPUX. I got around to reproducing this today,
and what I find is that the majority of the backend time is going into
simple scanning of the input statement:

Each sample counts as 0.01 seconds.
% cumulative self self total
time seconds seconds calls ms/call ms/call name
31.24 11.90 11.90 _mcount
19.51 19.33 7.43 10097 0.74 1.06 base_yylex
7.48 22.18 2.85 21953666 0.00 0.00 appendStringInfoChar
5.88 24.42 2.24 776 2.89 2.89 pglz_compress
4.36 26.08 1.66 21954441 0.00 0.00 pq_getbyte
3.57 27.44 1.36 7852141 0.00 0.00 addlit
3.26 28.68 1.24 1552 0.80 0.81 scanstr
2.84 29.76 1.08 779 1.39 7.18 pq_getstring
2.31 30.64 0.88 10171 0.09 0.09 _doprnt
2.26 31.50 0.86 776 1.11 1.11 byteain
2.07 32.29 0.79 msquadloop
1.60 32.90 0.61 7931430 0.00 0.00 memcpy
1.18 33.35 0.45 chunks
1.08 33.76 0.41 46160 0.01 0.01 strlen
1.08 34.17 0.41 encore
1.05 34.57 0.40 8541 0.05 0.05 XLogInsert
0.89 34.91 0.34 appendStringInfo

60% of the call graph time is accounted for by these two areas:

index % time self children called name
7.43 3.32 10097/10097 yylex [14]
[13] 41.0 7.43 3.32 10097 base_yylex [13]
1.36 0.61 7852141/7852141 addlit [28]
1.24 0.01 1552/1552 scanstr [30]
0.02 0.03 3108/3108 ScanKeywordLookup [99]
0.00 0.02 2335/2335 yy_get_next_buffer [144]
0.02 0.00 776/781 strtol [155]
0.00 0.01 777/3920 MemoryContextStrdup [108]
0.00 0.00 1/1 base_yy_create_buffer [560]
0.00 0.00 4675/17091 isupper [617]
0.00 0.00 1556/1556 yy_get_previous_state [671]
0.00 0.00 779/779 yywrap [706]
0.00 0.00 1/2337 base_yy_load_buffer_state [654]
-----------------------------------------------
1.08 4.51 779/779 pq_getstr [17]
[18] 21.4 1.08 4.51 779 pq_getstring [18]
2.85 0.00 21953662/21953666 appendStringInfoChar [20]
1.66 0.00 21954441/21954441 pq_getbyte [29]
-----------------------------------------------

While we could probably do a little bit to speed up pg_getstring and its
children, it's not clear that we can do anything about yylex, which is
flex output code not handmade code, and is probably well-tuned already.

Bottom line: feeding huge strings through the lexer is slow.

regards, tom lane

Tom,

Thanks for looking into this. What you found is about what I expected.
What this means to me is that there should be a new todo item along
the lines:
* allow binding query args over FE/BE protocol

Generally what makes the statement large is the values being
inserted/updated or being specified in the where clause. If it were
possible from the client declare a cursor with bind variables (i.e. $1,
$2, ...), then later bind in the values and execute the cursor, it
should be possible to work around this problem.

For example the Oracle SQL*Net protocol (Oracle's equivalent to
postgres' FE/BE) has seven distinct operations that can be done on a
cursor from the client: open, bind, describe, parse, execute, fetch,
close. This perhaps is a little bit more control than you need, but you
get the general idea.

Currently the FE/BE protocol has two modes for executing queries: 1)
send the entire query as a string and get back the entire result set in
one go, or 2) use cursors in sql, which still sends the entire query in
one string, but allows you to get back a subset of the results. It
would be nice to have even more control than either of these two options
offer.

It seems to me that the server has all the functionality to do this,
because plpgsql supports it in 7.2. It just isn't can't be done from
the client.

I have been contemplating creating some SPI functions that could be
called from the client that would implement this type of functionality
and then enhancing the JDBC driver to use them instead of the regular
query execution of the FE/BE protocol. (i.e. one function to "declare"
a cursor with $1 placeholders, a second function to bind values and
execute the cursor, and then using standard fetch sql statement to get
results, therefore it could be done without changing the actual
protocol). This could also support the client explicitly "caching"
commonly used cursors and just rebinding/reexecuting them to avoid
having to reparse/replan commonly used queries.

If this all isn't too off the wall, I would like to see something along
these lines added to the todo list and perhapes this email thread along
with. At a minimum I would like to hear others opinions on the subject.

thanks,
--Barry

Tom Lane wrote:

> Barry Lind <barry(at)xythos(dot)com> writes:
>
>>In looking at some performance issues (I was trying to look at the
>>overhead of toast) I found that large insert statements were very slow.
>>...
>>I tried to run gprof to see where the time was going, but for some
>>reason the gprof output on my gmon.out file doesn't have any timing
>>information (all times are reported as 0.0) and I haven't been able to
>>figure out why yet.
>>
>
> That seems to be a common disease among Linuxen; dunno why. gprof
> works fine for me on HPUX. I got around to reproducing this today,
> and what I find is that the majority of the backend time is going into
> simple scanning of the input statement:
>
> Each sample counts as 0.01 seconds.
> % cumulative self self total
> time seconds seconds calls ms/call ms/call name
> 31.24 11.90 11.90 _mcount
> 19.51 19.33 7.43 10097 0.74 1.06 base_yylex
> 7.48 22.18 2.85 21953666 0.00 0.00 appendStringInfoChar
> 5.88 24.42 2.24 776 2.89 2.89 pglz_compress
> 4.36 26.08 1.66 21954441 0.00 0.00 pq_getbyte
> 3.57 27.44 1.36 7852141 0.00 0.00 addlit
> 3.26 28.68 1.24 1552 0.80 0.81 scanstr
> 2.84 29.76 1.08 779 1.39 7.18 pq_getstring
> 2.31 30.64 0.88 10171 0.09 0.09 _doprnt
> 2.26 31.50 0.86 776 1.11 1.11 byteain
> 2.07 32.29 0.79 msquadloop
> 1.60 32.90 0.61 7931430 0.00 0.00 memcpy
> 1.18 33.35 0.45 chunks
> 1.08 33.76 0.41 46160 0.01 0.01 strlen
> 1.08 34.17 0.41 encore
> 1.05 34.57 0.40 8541 0.05 0.05 XLogInsert
> 0.89 34.91 0.34 appendStringInfo
>
> 60% of the call graph time is accounted for by these two areas:
>
> index % time self children called name
> 7.43 3.32 10097/10097 yylex [14]
> [13] 41.0 7.43 3.32 10097 base_yylex [13]
> 1.36 0.61 7852141/7852141 addlit [28]
> 1.24 0.01 1552/1552 scanstr [30]
> 0.02 0.03 3108/3108 ScanKeywordLookup [99]
> 0.00 0.02 2335/2335 yy_get_next_buffer [144]
> 0.02 0.00 776/781 strtol [155]
> 0.00 0.01 777/3920 MemoryContextStrdup [108]
> 0.00 0.00 1/1 base_yy_create_buffer [560]
> 0.00 0.00 4675/17091 isupper [617]
> 0.00 0.00 1556/1556 yy_get_previous_state [671]
> 0.00 0.00 779/779 yywrap [706]
> 0.00 0.00 1/2337 base_yy_load_buffer_state [654]
> -----------------------------------------------
> 1.08 4.51 779/779 pq_getstr [17]
> [18] 21.4 1.08 4.51 779 pq_getstring [18]
> 2.85 0.00 21953662/21953666 appendStringInfoChar [20]
> 1.66 0.00 21954441/21954441 pq_getbyte [29]
> -----------------------------------------------
>
> While we could probably do a little bit to speed up pg_getstring and its
> children, it's not clear that we can do anything about yylex, which is
> flex output code not handmade code, and is probably well-tuned already.
>
> Bottom line: feeding huge strings through the lexer is slow.
>
> regards, tom lane
>
>

Barry Lind <barry(at)xythos(dot)com> writes:
> It seems to me that the server has all the functionality to do this,
> because plpgsql supports it in 7.2. It just isn't can't be done from
> the client.

True. Basically it's a shortcoming in the FE/BE protocol. There's
some other work that'd have to be done, but extending the protocol
would be the main bit.

> I have been contemplating creating some SPI functions that could be
> called from the client that would implement this type of functionality
> and then enhancing the JDBC driver to use them instead of the regular
> query execution of the FE/BE protocol.

Interesting as a proof-of-concept hack, but I'd sure not want to see
it shipped as a production solution. For one thing, the existing
fastpath protocol is itself too broken to encourage more widespread
use of. (See the various comments in fastpath.c.)

regards, tom lane

> Tom,
>
> Thanks for looking into this. What you found is about what I expected.
> What this means to me is that there should be a new todo item along
> the lines:
> * allow binding query args over FE/BE protocol

OK, I have added this to the TODO list in the cache section because it
is really just a special case of cached query plans where some constants
change. I know it is a litte more than that, but the basic issues are
the same.

--
Bruce Momjian | http://candle.pha.pa.us
pgman(at)candle(dot)pha(dot)pa(dot)us | (610) 853-3000
+ If your life is a hard drive, | 830 Blythe Avenue
+ Christ can be your backup. | Drexel Hill, Pennsylvania 19026