The Question: Why MySQL performance goes down for queries joining nearly empty tables when executed in parallel?
Below is more detailed explanation of the issue I'm facing. I have two tables in MySQL
CREATE TABLE first (
num int(10) NOT NULL,
UNIQUE KEY key_num (num)
) ENGINE=InnoDB
CREATE TABLE second (
num int(10) NOT NULL,
num2 int(10) NOT NULL,
UNIQUE KEY key_num (num, num2)
) ENGINE=InnoDB
The first one contains about a thousand records. The second one is empty or contains a very few records. It also contains double index which somehow relates to the issue: the problem goes away for single index. Now I'm trying to make a lot of identical queries to those tables in parallel. Each query looks like this:
SELECT first.num
FROM first
LEFT JOIN second AS second_1 ON second_1.num = -1 # non-existent key
LEFT JOIN second AS second_2 ON second_2.num = -2 # non-existent key
LEFT JOIN second AS second_3 ON second_3.num = -3 # non-existent key
LEFT JOIN second AS second_4 ON second_4.num = -4 # non-existent key
LEFT JOIN second AS second_5 ON second_5.num = -5 # non-existent key
LEFT JOIN second AS second_6 ON second_6.num = -6 # non-existent key
WHERE second_1.num IS NULL
AND second_2.num IS NULL
AND second_3.num IS NULL
AND second_4.num IS NULL
AND second_5.num IS NULL
AND second_6.num IS NULL
The issue I'm getting is that instead of having a nearly linear raise in performance on 8 core machine I actually have a drop. Namely having one process, the typical number of requests per second I have is about 200. Having two processes instead of expected increase up to 300 - 400 queries per second I actually have a drop down to 150. For 10 processes I have only 70 queries per seconds. The Perl code I'm using for testing is shown below:
#!/usr/bin/perl
use strict;
use warnings;
use DBI;
use Parallel::Benchmark;
use SQL::Abstract;
use SQL::Abstract::Plugin::InsertMulti;
my $children_dbh;
foreach my $second_table_row_count (0, 1, 1000) {
print '#' x 80, "\nsecond_table_row_count = $second_table_row_count\n";
create_and_fill_tables(1000, $second_table_row_count);
foreach my $concurrency (1, 2, 3, 4, 6, 8, 10, 20) {
my $bm = Parallel::Benchmark->new(
'benchmark' => sub {
_run_sql();
return 1;
},
'concurrency' => $concurrency,
'time' => 3,
);
my $result = $bm->run();
}
}
sub create_and_fill_tables {
my ($first_table_row_count, $second_table_row_count) = @_;
my $dbh = dbi_connect();
{
$dbh->do(q{DROP TABLE IF EXISTS first});
$dbh->do(q{
CREATE TABLE first (
num int(10) NOT NULL,
UNIQUE KEY key_num (num)
) ENGINE=InnoDB
});
if ($first_table_row_count) {
my ($stmt, @bind) = SQL::Abstract->new()->insert_multi(
'first',
['num'],
[map {[$_]} 1 .. $first_table_row_count],
);
$dbh->do($stmt, undef, @bind);
}
}
{
$dbh->do(q{DROP TABLE IF EXISTS second});
$dbh->do(q{
CREATE TABLE second (
num int(10) NOT NULL,
num2 int(10) NOT NULL,
UNIQUE KEY key_num (num, num2)
) ENGINE=InnoDB
});
if ($second_table_row_count) {
my ($stmt, @bind) = SQL::Abstract->new()->insert_multi(
'second',
['num'],
[map {[$_]} 1 .. $second_table_row_count],
);
$dbh->do($stmt, undef, @bind);
}
}
}
sub _run_sql {
$children_dbh ||= dbi_connect();
$children_dbh->selectall_arrayref(q{
SELECT first.num
FROM first
LEFT JOIN second AS second_1 ON second_1.num = -1
LEFT JOIN second AS second_2 ON second_2.num = -2
LEFT JOIN second AS second_3 ON second_3.num = -3
LEFT JOIN second AS second_4 ON second_4.num = -4
LEFT JOIN second AS second_5 ON second_5.num = -5
LEFT JOIN second AS second_6 ON second_6.num = -6
WHERE second_1.num IS NULL
AND second_2.num IS NULL
AND second_3.num IS NULL
AND second_4.num IS NULL
AND second_5.num IS NULL
AND second_6.num IS NULL
});
}
sub dbi_connect {
return DBI->connect(
'dbi:mysql:'
. 'database=tmp'
. ';host=localhost'
. ';port=3306',
'root',
'',
);
}
And for compare queries like this executed in concurrent with increasing performance:
SELECT first.num
FROM first
LEFT JOIN second AS second_1 ON second_1.num = 1 # existent key
LEFT JOIN second AS second_2 ON second_2.num = 2 # existent key
LEFT JOIN second AS second_3 ON second_3.num = 3 # existent key
LEFT JOIN second AS second_4 ON second_4.num = 4 # existent key
LEFT JOIN second AS second_5 ON second_5.num = 5 # existent key
LEFT JOIN second AS second_6 ON second_6.num = 6 # existent key
WHERE second_1.num IS NOT NULL
AND second_2.num IS NOT NULL
AND second_3.num IS NOT NULL
AND second_4.num IS NOT NULL
AND second_5.num IS NOT NULL
AND second_6.num IS NOT NULL
Testing results, cpu and disk usage measurements are here:
* table `first` have 1000 rows * table `second` have 6 rows: `[1,1],[2,2],..[6,6]` For query: SELECT first.num FROM first LEFT JOIN second AS second_1 ON second_1.num = -1 # non-existent key LEFT JOIN second AS second_2 ON second_2.num = -2 # non-existent key LEFT JOIN second AS second_3 ON second_3.num = -3 # non-existent key LEFT JOIN second AS second_4 ON second_4.num = -4 # non-existent key LEFT JOIN second AS second_5 ON second_5.num = -5 # non-existent key LEFT JOIN second AS second_6 ON second_6.num = -6 # non-existent key WHERE second_1.num IS NULL AND second_2.num IS NULL AND second_3.num IS NULL AND second_4.num IS NULL AND second_5.num IS NULL AND second_6.num IS NULL Results: concurrency: 1, speed: 162.910 / sec concurrency: 2, speed: 137.818 / sec concurrency: 3, speed: 130.728 / sec concurrency: 4, speed: 107.387 / sec concurrency: 6, speed: 90.513 / sec concurrency: 8, speed: 80.445 / sec concurrency: 10, speed: 80.381 / sec concurrency: 20, speed: 84.069 / sec System usage after for last 60 minutes of running query in 6 processes: $ iostat -cdkx 60 avg-cpu: %user %nice %system %iowait %steal %idle 74.82 0.00 0.08 0.00 0.08 25.02 Device: rrqm/s wrqm/s r/s w/s rkB/s wkB/s avgrq-sz avgqu-sz await svctm %util sda1 0.00 0.00 0.00 0.12 0.00 0.80 13.71 0.00 1.43 1.43 0.02 sdf10 0.00 0.00 0.00 0.03 0.00 1.07 64.00 0.00 10.00 5.00 0.02 sdf4 0.00 0.00 0.00 0.03 0.00 1.07 64.00 0.00 30.00 15.00 0.05 sdm 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 sdf8 0.00 0.00 0.00 0.37 0.00 1.24 6.77 0.00 5.00 3.18 0.12 sdf6 0.00 0.00 0.00 0.03 0.00 1.07 64.00 0.00 10.00 5.00 0.02 sdf9 0.00 0.00 0.00 0.03 0.00 1.07 64.00 0.00 0.00 0.00 0.00 sdf 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 sdf3 0.00 0.00 0.00 0.08 0.00 1.33 32.00 0.00 4.00 4.00 0.03 sdf2 0.00 0.00 0.00 0.17 0.00 1.37 16.50 0.00 3.00 3.00 0.05 sdf15 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 sdf14 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 sdf1 0.00 0.00 0.00 0.05 0.00 0.40 16.00 0.00 0.00 0.00 0.00 sdf13 0.00 0.00 0.00 0.03 0.00 1.07 64.00 0.00 10.00 5.00 0.02 sdf5 0.00 0.00 0.00 0.03 0.00 1.07 64.00 0.00 50.00 25.00 0.08 sdm2 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 sdm1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 sdf12 0.00 0.00 0.00 0.03 0.00 1.07 64.00 0.00 10.00 5.00 0.02 sdf11 0.00 0.00 0.00 0.03 0.00 1.07 64.00 0.00 10.00 5.00 0.02 sdf7 0.00 0.00 0.00 0.03 0.00 1.07 64.00 0.00 10.00 5.00 0.02 md0 0.00 0.00 0.00 0.97 0.00 13.95 28.86 0.00 0.00 0.00 0.00 ################################################################################ For query: SELECT first.num FROM first LEFT JOIN second AS second_1 ON second_1.num = 1 # existent key LEFT JOIN second AS second_2 ON second_2.num = 2 # existent key LEFT JOIN second AS second_3 ON second_3.num = 3 # existent key LEFT JOIN second AS second_4 ON second_4.num = 4 # existent key LEFT JOIN second AS second_5 ON second_5.num = 5 # existent key LEFT JOIN second AS second_6 ON second_6.num = 6 # existent key WHERE second_1.num IS NOT NULL AND second_2.num IS NOT NULL AND second_3.num IS NOT NULL AND second_4.num IS NOT NULL AND second_5.num IS NOT NULL AND second_6.num IS NOT NULL Results: concurrency: 1, speed: 875.973 / sec concurrency: 2, speed: 944.986 / sec concurrency: 3, speed: 1256.072 / sec concurrency: 4, speed: 1401.657 / sec concurrency: 6, speed: 1354.351 / sec concurrency: 8, speed: 1110.100 / sec concurrency: 10, speed: 1145.251 / sec concurrency: 20, speed: 1142.514 / sec System usage after for last 60 minutes of running query in 6 processes: $ iostat -cdkx 60 avg-cpu: %user %nice %system %iowait %steal %idle 74.40 0.00 0.53 0.00 0.06 25.01 Device: rrqm/s wrqm/s r/s w/s rkB/s wkB/s avgrq-sz avgqu-sz await svctm %util sda1 0.00 0.00 0.00 0.02 0.00 0.13 16.00 0.00 0.00 0.00 0.00 sdf10 0.00 0.00 0.00 0.03 0.00 1.07 64.00 0.00 10.00 5.00 0.02 sdf4 0.00 0.00 0.00 0.03 0.00 1.07 64.00 0.00 10.00 5.00 0.02 sdm 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 sdf8 0.00 0.00 0.00 0.03 0.00 1.07 64.00 0.00 10.00 5.00 0.02 sdf6 0.00 0.00 0.00 0.03 0.00 1.07 64.00 0.00 0.00 0.00 0.00 sdf9 0.00 0.00 0.00 0.03 0.00 1.07 64.00 0.00 10.00 5.00 0.02 sdf 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 sdf3 0.00 0.00 0.00 0.13 0.00 2.67 40.00 0.00 3.75 2.50 0.03 sdf2 0.00 0.00 0.00 0.23 0.00 2.72 23.29 0.00 2.14 1.43 0.03 sdf15 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 sdf14 0.00 0.00 0.00 0.98 0.00 0.54 1.10 0.00 2.71 2.71 0.27 sdf1 0.00 0.00 0.00 0.08 0.00 1.47 35.20 0.00 8.00 6.00 0.05 sdf13 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 sdf5 0.00 0.00 0.00 0.03 0.00 1.07 64.00 0.00 10.00 5.00 0.02 sdm2 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 sdm1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 sdf12 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 sdf11 0.00 0.00 0.00 0.03 0.00 1.07 64.00 0.00 0.00 0.00 0.00 sdf7 0.00 0.00 0.00 0.03 0.00 1.07 64.00 0.00 10.00 5.00 0.02 md0 0.00 0.00 0.00 1.70 0.00 15.92 18.74 0.00 0.00 0.00 0.00 ################################################################################ And this server has lots of free memory. Example of top: top - 19:02:59 up 4:23, 4 users, load average: 4.43, 3.03, 2.01 Tasks: 218 total, 1 running, 217 sleeping, 0 stopped, 0 zombie Cpu(s): 72.8%us, 0.7%sy, 0.0%ni, 26.3%id, 0.0%wa, 0.0%hi, 0.0%si, 0.1%st Mem: 71701416k total, 22183980k used, 49517436k free, 284k buffers Swap: 0k total, 0k used, 0k free, 1282768k cached PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 2506 mysql 20 0 51.7g 17g 5920 S 590 25.8 213:15.12 mysqld 9348 topadver 20 0 72256 11m 1428 S 2 0.0 0:01.45 perl 9349 topadver 20 0 72256 11m 1428 S 2 0.0 0:01.44 perl 9350 topadver 20 0 72256 11m 1428 S 2 0.0 0:01.45 perl 9351 topadver 20 0 72256 11m 1428 S 1 0.0 0:01.44 perl 9352 topadver 20 0 72256 11m 1428 S 1 0.0 0:01.44 perl 9353 topadver 20 0 72256 11m 1428 S 1 0.0 0:01.44 perl 9346 topadver 20 0 19340 1504 1064 R 0 0.0 0:01.89 top
Does anyone have an idea why performance decreased for query with non-existent keys?