Today we'll cover:

• What is profiling and why do we need it?
• How to profile?
• How to view and analyze profiling results?
• What are common bottlenecks in our code?
• How to maintain performance?

Why profile if application works fast?

Fast?

How fast?

0.5 second — is it fast enough?

«Everything is relative»

Albert Einstein

Why optimizing?

• faster response
• more RPS
• save resources → money

Profiling

— is the process of application runtime characteristics measurement for performance analysis and optimization

Profilers are a source of knowledge on:

• where
• on which operations
• and how much

resources are spent during an application run

Ha, I don't need profiler to know where the bottleneck is.

I'll fix it right now.

Nope. Probably nope. The whole application optimization without profiling is a shooting in the dark.

What application parameters might be of interest?

• execution time
• CPU time
• memory usage
• disk operations time (IO)
• network time
• DBMS response time*

How it works?

• Profiler is a special C app, compiled as PHP extension
• Runs on demand or for every request and records metrics necessary.
• Records data in a profiling log

How to profile in PHP?

• PHP itself
• XDebug
• XHprof
• Blackfire

PHP itself

  $start = microtime(true);
  $this->check();
  $time = microtime(true) - $start;
                    

XDebug

• xdebug.org
• docs

Not a profiler.

Debugger with a profile function

Why not?

• huge overhead (app execution time × 2+)
• optimization for XDebug is 90% useless for when it gets disabled
• not for production server

But if you still debug with print...

Must have for debug!

Supported by:

• PHPStorm
• NetBeans
• Atom
• Sublime Text
• Cloud9
• vim
• ...

Step-by-step debug

Stacktrace

View & modify variables in scope

XHprof

• XHprof on PECL
• Docs on PHP.net

Advantages:

• minimal overhead (~10%)
• may be used on production

Disadvantages:

• does not support PHP 7 officially.
  Forks do:

  • Tideways.io
  • RustJason/xhprof
• can measure execution and CPU time and memory, but usually it's enough

Easy to start!

Configure your php.ini

[xhprof]
extension=xhprof.so
xhprof.output_dir="/tmp/xhprof"

Enable in application entry script

<?php

xhprof_enable(XHPROF_FLAGS_CPU | XHPROF_FLAGS_MEMORY);
(new Application())->run();
$xhprof_data = xhprof_disable();

xhprof_disable() returns array of calls

["yii\base\Module::runAction==>yii\base\Controller::runAction"] =>
    array(5) {
        ["ct"] => int(1)
        ["wt"] => int(6718)
        ["cpu"] => int(6720)
        ["mu"] => int(617624)
        ["pmu"] => int(672904)
    }

• ct – called times
• wt – wall time, microseconds
• cpu – CPU time, microseconds
• mu – memory usage, bytes
• pmu – peak memory usage, bytes

So what's next?

Save!

    <?php
    xhprof_enable(XHPROF_FLAGS_CPU | XHPROF_FLAGS_MEMORY);
    (new Application())->run();
    $xhprof_data = xhprof_disable();

    $XHPROF_LIB_ROOT = "/opt/xhprof/xhprof_lib";
    include_once $XHPROF_LIB_ROOT . "/utils/xhprof_lib.php";
    include_once $XHPROF_LIB_ROOT . "/utils/xhprof_runs.php";

    $xhprof_runs = new XHProfRuns_Default();
    $run_id = $xhprof_runs->save_run($xhprof_data, 'my_application');

These PHP files are available in XHprof repository. You can either require them or put them in the project. You can even modify them.

How to check results?

Take xhprof_lib and xhprof_html directories from the same repo, and start a web-server with root in xhprof_html.

Knowing profile IDs you can compare two of them:

http://xhprof.localhost/index.php?run1=__ID1__&run2=__ID2__

Em... Not cool.

Is there something fancy?

Ofc! Get graphviz:

apt-get install graphviz

Then click [View Full Callgraph] on the profile page and get...

Keep calm and read:

• Class and method name
• Execution time including calls from this method
• Execution time excluding calls from this method
• Method calls number

Can we get even more information?

Drawer uses PHP and it's OpenSource so you can easily add CPU, RAM usage or whatever.

Blackfire

• blackfire.io
• Docs

• Started as XHprof fork, but was rewritten
• Is a product of Sensio Labs
• Small overhead
• Can be used on production servers
• Linux, FreeBSD, Windows, MacOS support
• PHP 7 support
• It's freemium. Wall time and memory usage for free. CPU, IO, network and other features starts from 19,90€
• Quite good documentation

How to install?

• Described very well in the manual

Blackfire consists of

• Website to view and analyze saved profiles
• PHP extension
• Agent to send saved profiles to the website
• Chrome extension
• CLI tools

Turns on by trigger

Profiles comparision

How had I reached 10x speedup?

Disabled xDebug

Blackfire has many interesting features:

• Manual XHprof profiles upload
• Metrics testing in PHPUnit
• Build profiling in Travis or Jenkins
• Notifications on GitHub, Gitlab, Slack, HipChat, e-mail
• Average profile of N executions

  blackfire --samples 10 run php my-script.php

Now we know how to draw, but how to make the app faster?

Analyze!

What should we profile first?

• Things that are slow
• Application initialization
• Popular pages

Profile analyzing

• Follow the hot path
• Think what causes the result
• Find the right method to blame
• Fix it
• ???
• PROFIT!

Popular bottlenecks

• Bad/complex algorithms
• Nested loops
• Disk IO
• Bad SQL queries
• Communication with the Universe
• Console commands execution

How to deal with it?

Bad/complex algorithms

• Profile, analyze, optimize
• Look for a different solution
• Make async if possible
• Cache

Disk IO

• Lazy loading
• SSD
• Tune realpath cache
• Tune opcache
• Use faster storages (Memcache, Redis)

Bad SQL queries

• Analyze query EXPLAIN and optimize
• Configure indexes
• Tune your DBMS
• Avoid complex queries

Communication with the Universe

• SSL handshake takes much
• Be prepared, your endpoint will get down
• Async where possible (AMQP, RabbitMQ)

Console commands execution

• Async where possible
• Minimize
• Cache

Time for a small example

Checking the SQL query

SELECT  *
FROM    sale
WHERE   date > '2017-04-01'::datetime
    AND date < '2017-05-01'::datetime;

• No indices?
• To much data to return?
• Slow network connection to DBMS?
• Table is locked too often?
• Combination of the above?
• Everything above is wrong?

Profilers on production

• Use triggers to activate
• Remember, that profiles take disk space

Frameworks' help in application performance

• Typical problems are already solved
• Inspiration for a good code and/or performance tricks
• Debug panel

Debug panel

Application logs

SQL queries logs

Fast profiling

    Yii::beginProfile('Checking my new algo');
    $system->compute();
    Yii::endProfile('Checking my new algo');
                    

We should try to:

• Remember about architecture importance and clean code
• Avoid god classes and methods
• Remember about algorithms complexity
• Don't spend too much time on early optimization

And don't forget to:

• Upgrade to PHP 7 (up to +40%) or even PHP 7.1 (up to 30% on top)
• Configure your caches (Opcache, Realpath cache)
• Think about what you do :)