1 Million SQL Queries per second: GA MariaDB 10.1 on POWER8

A couple of days ago, MariaDB announced that MariaDB 10.1 is stable GA – around 19 months since the GA of MariaDB 10.0. With MariaDB 10.1 comes some important scalabiity improvements, especially for POWER8 systems. On POWER, we’re a bit unique in that we’re on the higher end of CPUs, have many cores, and up to 8 threads per core (selectable at runtime: 1, 2, 4 or 8/core) – so a dual socket system can easily be a 160 thread machine.

Recently, we (being IBM) announced availability of a couple of new POWER8 machines – machines designed for Linux and cloud environments. They are very much OpenPower machines, and more info is available here: http://www.ibm.com/marketplace/cloud/commercial-computing/us/en-us

Combine these two together, with Axel Schwenke running some benchmarks and you get 1 Million SQL Queries per second with MariaDB 10.1 on POWER8.

Having worked a lot on both MySQL for POWER and the firmware that ships in the S882LC, I’m rather happy that 1 Million queries per second is beyond what it was in June 2014, which was a neat hack on MySQL 5.7 that showed the potential of MySQL on POWER8 but wasn’t yet a product. Now, you can run a GA release of MariaDB on GA POWER8 hardware designed for scale-out cloud environments and get 1 Million SQL queries/second (with fewer cores than my initial benchmark last year!)

What’s even more impressive is that this million queries per second is in a KVM guest!

1 million SQL Queries Per Second: MySQL 5.7 on POWER8

I’ve previously covered MySQL 5.6 on POWER (with patch), MySQL 5.6 Performance on POWER8 (spoiler: new performance record) and MySQL 5.7 on POWER.

Of course, The postings on this site are my own and don’t necessarily represent IBM’s positions, strategies or opinions. Also, these numbers should be considered preliminary, but trust me – I did get them and it’s not April 1st.

From my last post, you saw that with my preliminary patch for MySQL 5.7 to work on POWER, we could easily match the previous record for sysbench point select queries per second (i.e. key lookups). In fact, we could exceed the published record by a little bit which is kind of nice. At around 630kQPS, one could be rather happy.

But we still had 30-40% idle CPU on POWER8. This led me to file the following bug report:

  • Bug 72829: LOCK_grant is major contention point, leaves 30-40% idle CPU.

What’s going on is that there’s a rwlock in the MySQL Server that ensures that writers don’t collide with readers to the data structures describing the GRANTs (i.e. who has access to what). If you run a GRANT statement, it gets a writer lock, and nobody can read (i.e. check permissions) while everything is being updated. If you run a normal SQL statement, you get a read lock (non-exclusive) and can check permissions appropriately.

It’s been known for a long time that LOCK_grant was a bottleneck. Typically, some people have run with skip-grant-tables to help shorten the time the lock as held (as in MySQL you still take the mutex even though you’ve started the server with skip-grant-tables).

In Drizzle, we fixed that – moving authentication and authorization completely behind plugin APIs and if you didn’t load plugins for them, you executed near enough to zero instructions that it didn’t matter.

In my experiments, enabling skip-grant-tables actually hurt performance rather than helped. More investigation is needed, but it seems that simply the act of acquiring and releasing the rdlock is now a major bottleneck in some benchmarks (such as sysbench point select).

It turns out that this is a well known problem in other pieces of software (e.g. Linux kernel) and is pretty much what RCU (Read Copy Update) is best at. As far back as 2006 I remember attempting to get my head around RCU so that one day we could use it in MySQL or MySQL Cluster.

Another simpler method is simply splitting the mutex, with readers able to acquire any one of N mutexes and writers needing to acquire them all. This penalizes writers, but unless you’re executing a lot of GRANTs, you’re probably safe.

So… what is the theoretical maximum performance if this bottleneck went away?

I wrote a quick patch that just commented out the rdlock acquisition of LOCK_grant in the hot codepath of sysbench point selects. I wasn’t running GRANT statements at runtime so this was “safe”.

This patch is not production ready, it’s merely useful for demonstrating where we could be with MySQL 5.7 on POWER8 if one last bottleneck is fixed.

My results? Slightly over ONE MILLION QUERIES PER SECOND!

This is roughly twice the previous record.

This is with a dual socket 24 core POWER8 with SMT8 and DSCR=1 on 8 tables with sysbench 0.4.8. Sysbench itself is using a non-trivial amount of CPU and I could probably decently beat this number if I rewrote sysbench using the nonblocking API in libdrizzle (back when me made the Drizzle performance regression tests use a libdrizzle-ified sysbench we got double digit percentage improvement in our sysbench numbers).

There’s still around 7-10% idle CPU time… so there’s more room to grow.

Lacking a physical gauntlet to throw down, I’ll just have to submit a conference paper somewhere so that I can do that in person.

I really hope that we’re able to fix this bottleneck in MySQL 5.7 so that MySQL 5.7 will ship being able to do over a million queries per second. From SQL.

Finding out What’s Next at BarCampMel 2012 with Drizzle, SQL, JavaScript and a web browser

Just for the pure insane fun of it, I accepted the challenge of “what can you do with the text format of the schedule?” for BarCampMel. I’m a database guy, so I wanted to load it into a database (which would be Drizzle), and I wanted it to be easy to keep it up to date (this is an unconference after all).

So… the text file itself isn’t in any standard format, so I’d have to parse it. I’m lazy and didn’t want to leave the comfort of the database. Luckily, inside Drizzle, we have a js plugin that lets you execute arbitrary JavaScript. Parsing solved. I needed to get the program and luckily we have the http_functions plugin that uses libcurl to allow us to perform HTTP GET requests. I also wanted it in a table so I could query it when not online, so I needed to load the data. Luckily, in Drizzle we have the built in EXECUTE functionality, so I could just use the JavaScript to parse the response from the HTTP GET request and construct SQL to load the data into a table to then query.

So, grab your Drizzle server with “plugin-add=js” and “plugin-add=http_functions” in the config file or as options to drizzled (prefixed with –) and….

This simple one liner pulls the current schedule and puts it into a table called ‘schedule’:

SELECT EXECUTE(JS("function sql_quote(s) {return s ? '\"'+ s.replace('\"', '\\\"') + '\"' : 'NULL'} function DrizzleDateString(d) { function pad(n) { return n<10 ? '0'+n : n } return d.getFullYear()+'-'+pad(d.getMonth()+1)+'-'+pad(d.getDate())+' '+pad(d.getHours())+':'+pad(d.getMinutes())+':'+pad(d.getSeconds()) } var sql = 'COMMIT;CREATE TABLE IF NOT EXISTS schedule (start_time datetime, stage varchar(1000), mr2 varchar(1000), mr1 varchar(1000), duration int); begin; delete from schedule;' ; var time= new Date;var input= arguments[0].split(\"\\n\"); var entry = new Array(); var stage, mr2, mr1; for(var i=0; i < input.length; i++) { var p= input[i].match('^(.*?) (.*)$'); if(p) {if(p[1]=='Time') { time=new Date(Date.parse(p[2]));} if(p[1]=='Duration') { sql+='INSERT INTO schedule (start_time,stage,mr2,mr1,duration) VALUES (\"' + DrizzleDateString(time) + '\", ' + sql_quote(stage) + ', ' + sql_quote(mr2) + ',' + sql_quote(mr1) + ',' + p[2] + '); '; time= new Date(time.getTime()+p[2]*60*1000); stage= mr2= mr1= ''; } if(p[1]=='stage') {stage=p[2]} if (p[1]=='mr2') {mr2=p[2]} if (p[1]=='mr1') {mr1=p[2]} }}; sql+='COMMIT;'; sql", (select http_get('https://dl.dropbox.com/s/01yh7ji7pswjwwk/live-schedule.txt?dl=1'))));

Which you can then find out “what’s on now and coming up” with this query:

select * from schedule where start_time > DATE_ADD(now(), INTERVAL 9 HOUR) ORDER BY start_time limit 2\G
But it’s totally not fun having to jump to the command line all the time, and you may want it in JSON format for consuming with some web thing…. so you can load the json_server plugin and browse to the port that it’s running on (default 8086) and type the SQL in there and get a JSON response, or just look at the pretty table there.

Drizzle JSON interface merged


Currently a very early version of course, but it’s there in trunk if you want to play with it. Just have libcurl and libevent installed and you can submit queries via HTTP and JSON. Of course, the next steps are getting a true non-sql interface going and seeing how people go with it.


Here’s a nice challenge for you. What does the following do (or error out on?):

CREATE TABLE t1 (a int);
CREATE TABLE t2 (b int);

I’d be interested to know what a) you think it does and then b) if you were surprised when you went and typed it into your RDBMS of choice.

New CREATE TABLE performance record!

4 min 20 sec

So next time somebody complains about NDB taking a long time in CREATE TABLE, you’re welcome to point them to this :)

  • A single CREATE TABLE statement
  • It had ONE column
  • It was an ENUM column.
  • With 70,000 possible values.
  • It was 605kb of SQL.
  • It ran on Drizzle

This was to test if you could create an ENUM column with greater than 216 possible values (you’re not supposed to be able to) – bug 589031 has been filed.

How does it compare to MySQL? Well… there are other problems (Bug 54194 – ENUM limit of 65535 elements isn’t true filed). Since we don’t have any limitations in Drizzle due to the FRM file format, we actually get to execute the CREATE TABLE statement.

Still, why did this take four and a half minutes? I luckily managed to run poor man’s profiler during query execution. I very easily found out that I had this thread constantly running check_duplicates_in_interval(), which does a stupid linear search for duplicates. It turns out, that for 70,000 items, this takes approximately four minutes and 19.5 seconds. Bug 589055 CREATE TABLE with ENUM fields with large elements takes forever (where forever is defined as a bit over four minutes) filed.

So I replaced check_duplicates_in_interval() with a implementation using a hash table (boost::unordered_set actually) as I wasn’t quite immediately in the mood for ripping out all of TYPELIB from the server. I can now run the CREATE TABLE statement in less than half a second.

So now, I can run my test case in much less time and indeed check for correct behaviour rather quickly.

I do have an urge to find out how big I can get a valid table definition file to though…. should be over 32MB…

Embedded InnoDB: querying the configuration

I am rather excited about being able to do awesome things such as this to get the current configuration of your server:

    ->  WHERE NAME IN ("data_file_path", "data_home_dir");
| NAME           | VALUE |
| data_file_path | NULL  | 
| data_home_dir  | ./    | 
2 rows in set (0 sec)

    -> WHERE NAME IN ("data_file_path", "data_home_dir");
| NAME           | VALUE |
| data_file_path | NULL  | 
| data_home_dir  | ./    | 
2 rows in set (0 sec)

    -> WHERE NAME = "io_capacity";
| NAME        | VALUE |
| io_capacity | 200   | 
1 row in set (0 sec)

Coming soon: status in a table.

(this is for the upcoming embedded_innodb plugin, which using the API provided by Embedded InnoDB to implement a Storage Engine for Drizzle)

NDB$INFO with SQL hits beta

Bernhard blogged over at http://ocklin.blogspot.com/2010/02/mysql-cluster-711-is-there.html that MySQL Cluster 7.1.1 Beta has been released. The big feature (from my point of view) is the SQL interface on top of NDB$INFO. This means there is now full infrastructure from the NDB data nodes right out to SQL in the MySQL Server for adding monitoring to any bit of the internals of the data nodes.

Stewart learns SQL oddities…

What would you expect the following to fail with?

CREATE TABLE t1 (a int, b int);
insert into t1 values (100,100);
CREATE TEMPORARY TABLE t2 (a int, b int, primary key (a));
INSERT INTO t2 values(100,100);
CREATE TEMPORARY TABLE IF NOT EXISTS t2 (primary key (a)) select * from t1;

If you answered ER_DUP_ENTRY then you are correct.

From the manual:


If you use IF NOT EXISTS in a CREATE TABLE ... SELECT statement, any rows selected by the SELECT part are inserted regardless of whether the table already exists.

Does anybody else find this behaviour “interesting”?

online online online! (or restarts are for wusses)

I often see things go past my eyes where customers (and users – i.e. those that don’t send wads of cash our way and hence are not financially supporting my beer, curry and photography habits) have amazing uptime and reliability requirements.

When talking to businesses that use MySQL, it’s not uncommon to have the “if the DB is down, our business doesn’t operate” line bandied around. How people make sure this never happens can differ (hint: it often involves replication and good sysadmin practices).

One thing I like doing is making things easier for people. Sometimes it’s also a much more complicated problem than you’re initially led to believe.

I think configuration files are obsolete. Okay, maybe just for databases. Everything should be changable as an online operation. This should also be able to be done via a standard interface – in our case, SQL. This means it’s suddenly really easy to write portable UIs around the admin functionality (no getting the parsing and generation – most trickily, the modification of text based config files right) just the issuing of SQL to the server, relativly simple. This even enables web apps to tune the database a bit, opting for various amounts of automation for various applications – in a cross platform way!

One of my visions for NDB (MySQL Cluster) is to get rid of the (user visible) configuration file and manage everything through SQL (or management client, something like that). This way you could ALTER CLUSTER ADD NODE, ALTER CLUSTER SET DataMemory=4GB etc and things should “just work”, take however long it needs – without downtime.

In a clustered environment, we could do these operations transactionally so that in the event of node or system failure we have some hope of being in a nicely consistent state and that during system recovery (or node recovery) we’re not performing a configuration change in addition to restarting (e.g. if you edited a config file and then had a crash).

Config changes could also have EXPLAIN, a non-modifying operation that would EXPLAIN what would be done – e.g. rolling restart, taking approximately X minutes per node and Y minutes total. This could help in planning and scheduling of configuration changes.

(i wonder if that made any sense)