Add symfony

[waghanza]

Hi,

There is no php frameworks here.

The reason was that in php world, there is not appserver (something bridging between php source and http), so we have to use :

• php-fpm
• nginx

So more works than with the other framework, that's why it take so long.

This PR add symfony as the first php framework

Regards,

[aichholzer]

And you forgot to mention; php is sooo slow. 🐢

[OvermindDL1]

PHP is not really that slow anymore, especially once the bytecode cache is setup and all (which should be done on any properly set up PHP server).

[aichholzer]

@OvermindDL1 -Yep, still, all of that seems to be too much effort, IMHO. In any case, I'd be curious to see who well it performs against the other contenders.

@waghanza -I did not mean any offense with my previous comment. I apologize if it came across like that.

[OvermindDL1]

all of that seems to be too much effort

It's less effort than setting up a Ruby project by far at the very least and we have plenty of those here... ^.^;

[waghanza]

@aichholzer not at all :stuck_out_tongue: This framework shows me some weird results

Language (Runtime)	Framework (Middleware)	Requests / s	Latency	99 percentile	Throughput
node	rayo	82198.33	18325.00	273634.33	42.33 MB
php	symfony	40.00	2925631.00	6764337.33	0.14 MB

So I will adjust some params in the implementation

@OvermindDL1 between sinatra, laravel and flask (python, ruby and php) there is no very difference in term of efforts to create an app :stuck_out_tongue: except there is no appserver in php, something like puma or gunicorn

[waghanza]

@OvermindDL1 if you have any tip to optimize this php code, feel free to share

[waghanza]

@OvermindDL1 for example, using a socket connection between nginx and php-fpm (instead of a tcp connection) lead to other results

Language (Runtime)	Framework (Middleware)	Requests / s	Latency	99 percentile	Throughput
node	rayo	84144.33	14735.67	113253.67	41.51 MB
php	symfony	72971.33	146987.00	3547798.00	123.28 MB

[waghanza]

@fabpot @nicolas-grekas I found a huge throughput (php 7.1 and php 7.2) here on a basic symfony implementation on an API returning :

• an empty body on GET /
• the id on GET /user/:id
• an empty body on POST /user

could you checky my code https://github.com/waghanza/which_is_the_fastest/blob/php_symfony/php/symfony/src/Controller/ApplicationController.php ?

[nicolas-grekas]

@waghanza thanks for the ping. Can you create an issue on https://github.com/symfony/symfony/ so that someone from the community could have a look?

[stof]

@waghanza make sure you run this in the prod mode of Symfony, not in the debug mode. The debug mode is much slower (we trade speed in favor of developer experience to provide a lot of debugging tools about what happens)

[stof]

hmm, it seems to be set to prod and no-debug mode though

[waghanza]

@stof production mode is enabled, but no details about debug (if I refer to my knowledge ^^ of symfony, debug is disabled by default)

[OvermindDL1]

That latency on symfony implies to me that php is spooling up the script each time, no opcode cache set up yet? I can't think of any web hosting provider for PHP that doesn't use an opcode cache as without it php is essentially running in a debug kind of mode itself (reparsing and rerunning scripts each time)...

But yeah, I would expect PHP's latency to be higher each time anyway if using a framework like symfony. To have a low latency version you'd need a framework that is designed to keep running in memory (most aren't in PHP, that's a relatively recent creation in the PHP world).

[bonfante]

Hey, good work. Can I make a suggestion? Maybe changing the directory from 'symfony' to 'symfony_nginx'. This will allow others to add more setups such as httpd, swoole, or phpppm. For more performance other way to go is in config nginx and fpm with opcache.

[nicolas-grekas]

About the throughput, I suppose it depends on the size of the response, headers included? What's the typical response from the current setup and how does it differ from the other tested frameworks?

[waghanza]

@nicolas-grekas the througput is computed with wrk using the definition :

• an empty body on GET /
• the id on GET /user/:id
• an empty body on POST /user

The headers and else are left to the default (php-fpm / nginx). My results are weird, but I think due to what i am using (docker). I'll merge this here, and ask for advice after switching to a real-world usage (cloud or else).

Language (Runtime)	Framework (Middleware)	Requests / s	Latency	99 percentile	Throughput
ruby	rails	3828.33	16757.67	113968.67	3.68 MB
ruby	sinatra	16248.67	3941.67	46547.67	13.70 MB
ruby	roda	38744.33	1729.67	28286.67	12.30 MB
ruby	rack-routing	31118.00	2080.00	33025.00	5.79 MB
ruby	flame	19446.33	3308.33	37345.33	3.62 MB
node	express	48666.67	35435.00	620252.67	40.56 MB
node	fastify	66903.33	23119.00	367041.67	67.33 MB
node	polka	78586.33	16510.67	210428.67	39.21 MB
node	rayo	79417.67	20978.33	391976.67	38.50 MB
python	sanic	48544.67	19461.67	33328.00	28.16 MB
python	japronto	86920.00	11040.33	11565.33	35.31 MB
python	flask	19125.33	53313.33	167249.67	15.75 MB
python	django	11238.00	83441.67	209351.00	10.62 MB
python	tornado	1907.00	541693.00	3961058.33	1.39 MB
php	symfony	152.33	1117751.67	4560955.33	0.20 MB

[waghanza]

@Bonfante Your idea is awesome, however I think that having multiple implementations per framework here is counter-productive.

• Result complication
• More maintenance
• No real representation

For me this project is a representation of what it COULD be done, but as an average, for a more specific representation, we encourage to use this as a tool (eg : having your own implementations base on this project -> I will create some documentations for this)

[doanguyen]

Now I understand the whole picture from #421, and it's very interesting to me.

So @waghanza switch from tcp socket to unix socket for symfony, there are two things that I just found out:

1. unix socket is really fast, I tried to tweak the fastcgi_connect_timeout from nginx but it does not seem to work, and it makes Nginx fall down quickly then return 502 error (wrk unfortunately consider a successful request) instead of waiting until handle by PHP-FPM. TCP socket is slower but keeping the connection for a while, Nginx then needs to wait for the socket before accepting another connection. That plus the weird report from wrk makes the req/s from two setups is so different.
2. The abnormal throughput from symfony due to the fact that the response from successful requests are gzipped while 502 error is not.

[chiqui3d]

Hello @nicolas-grekas,

It doesn't seem right that Symfony should be so badly off in this Benchmark.

And I see that some things are missing:

• APP_DEBUG=0 and APP_SECRET is not defined
• The Kernel.php file has changed a lot since 3 years ago, plus this demo tries to load the services.yml file that does not exist, and is not configured with autowire.
• In public/index.php you are trying to get the environment variable SYMFONY_ENVand it is really APP_ENVand APP_DEBUG
• The boostrap.php file does not exist, although I think it is not necessary for this configuration.

[chiqui3d]

Here the tests come out more or less better https://www.techempower.com/benchmarks/#section=data-r19&hw=ph&test=fortune&l=zijzzz-1r

[chiqui3d]

And here other PHP tests for those who say that php is slow https://github.com/chiqui3d/php-faster-than-python-ruby-node

The slowest is Ruby