Web Framework Performance Comparison

This project provides representative performance measures across a wide field of web application frameworks. With much help from the community, coverage is quite broad and we are happy to broaden it further with contributions. The project presently includes frameworks on many languages including Go, Python, Java, Ruby, PHP, C#, Clojure, Groovy, Dart, JavaScript, Erlang, Haskell, Scala, Perl, Lua, C, and others. The current tests exercise plaintext responses, JSON seralization, database reads and writes via the object-relational mapper (ORM), collections, sorting, server-side templates, and XSS counter-measures. Future tests will exercise other components and greater computation.

Read more and see the results of our tests on Amazon EC2 and physical hardware. For descriptions of the test types that we run, see the test requirements section.

Join in the conversation at our Google Group, or chat with us on Freenode at #techempower-fwbm.

How do I contribute updates or new frameworks?

Setup a development environment, typically by using a remote server dedicated to this project or by using a virtual machine. Here's a short example:

# Clone the project, then move into the right directory
$ cd ~/FrameworkBenchmarks/deployment/vagrant-development
# Turn on the VM. Takes at least 20 minutes
$ vagrant up
# Enter the VM, then run a test
$ vagrant ssh
vagrant@TFB-all:~$ cd ~/FrameworkBenchmarks
vagrant@TFB-all:~/FrameworkBenchmarks$ toolset/run-tests.py --install server --mode verify --test beego

Then enable Travis-CI on your project fork, so that any commits you send to Github are automatically verified for correctness (e.g. meeting all benchmark requirements). This is as simple as going to travis-ci.org, using the Log in with Github button, and enabling Travis-CI for your fork. When your development is done and your changes pass the Travis-CI verification, submit a pull request with confidence that it can be merged quickly.

How do I run the benchmark myself?

If you plan to run the benchmark and compare results, you need to run in benchmark mode. We recommend having a minimum of three distinct computers with a fast network connection in between, all of which you should be comfortable installing a large amount of additional software on. One of these computers (the app server) must have passwordless SSH access to the other two (search Google for help, yielding references such as these: article 1 article 2 article 3), and on every computer you will need to have passwordless sudo access (Google for help). Once you have cloned our repository, run toolset/run-tests.py --help for detailed help on running in benchmark mode.

If you wish to benchmark on Amazon EC2, see our scripts for launching a benchmark-ready Amazon environment.

If you are not an expert, please ensure your setup can run in verify mode before attempting to run in benchmark mode.

Project Overview

The full benchmark requires at least three computers:

• app server: The computer that your framework will be launched on.
• load server: The computer that will generate client load. Also known as the client machine.
• database server: The computer that runs all the databases. Also knonw as the DB server.

This codebase (TechEmpower/FrameworkBenchmarks aka TFB) must be run on the app server. The codebase contains a number of framework directories, each of which contains one or more framework test implementations. While our current setup has many directories, we are working to consolidate related code into fewer directories
with more tests per directory.

When run, TFB will:

• select which framework tests are to be run based on command-line arguments you provide
• install the necessary software (both on the app server and other servers)
• launch the framework
• access the urls listed in the requirements and verify the responses
• launch the load generation software on the load server
• gather the results
• halt the framework

Configuration File Usage

TFB takes a large number of command-line arguments, and it can become tedious to specify them repeatedly. We recommend you create a benchmark.cfg file by copying the example benchmark.cfg.example file and modifying as needed. See toolset/run-tests.py --help for a description of each flag.

For running in verify mode, you can set the various hosts to localhost. For running in benchmark mode, you will need all of the servers' IP addresses.

Note: environment variables can also be used for a number of the arguments.

Installation Basics

After you have a configuration file, run the following to setup your various servers. We use the --install-only flag in our examples to prevent launching tests at this stage.

See here for additional details.

Setting up the load server

$ toolset/run-tests.py --install client --install-only

Setting up the database server

$ toolset/run-tests.py --install database --install-only

Setting up the app server

You can choose to selectively install components by using the --test and --exclude flags.

# Install just the software for beego (as an example)
$ toolset/run-tests.py --install server --test beego --verbose --install-only

# Install all php software but php-fuel (as another example)
$ toolset/run-tests.py --install server --test php* --exclude php-fuel --verbose --install-only

# Install *all* framework software. Expect this to take hours!
# If running on a remote server, use `screen` or `tmux` or `nohup` to 
# prevent the installation from being terminated if you are disconnected
$ toolset/run-tests.py --install server --verbose --install-only

Listing Tests

You can easily list all available tests

$ toolset/run-tests.py --list-tests
activeweb
activeweb-raw
aspnet
aspnet-jsonnet
aspnet-mongodb-raw
aspnet-mono
aspnet-mono-jsonnet
aspnet-mono-mongodb-raw
<snip>

Running Tests

There are a number of options that can be specified:

# Run a verification for test beego
$ toolset/run-tests.py --test beego --mode verify

# Run the default benchmark for the beego test
$ toolset/run-tests.py --test beego

# Specify which test types are run during benchmark
$ toolset/run-tests.py --test beego --type json
$ toolset/run-tests.py --test beego --type db
$ toolset/run-tests.py --test beego --type fortune

# Specify a number of options for how the load is generated
$ toolset/run-tests.py --test beego --max-concurrency 24 --max-threads 24 --duration 20 --max-queries 200

# Run a tiny benchmark
$ toolset/run-tests.py --test beego --max-threads 2 --max-concurrency 2 

Finding output logs

Logs file locations use the format results/ec2/latest/logs/wt/err.txt. The general structure is results/<run name>/<timestamp>/logs/<test name>/<file> You can use the --name flag to change the <run name> If you re-run the same test multiple times, you will get a different folder for each <timestamp>, although the latest folder will be kept up to date. The <test name> is simply the name of the test type you ran, and <file> is either out.txt or err.txt (these are the logout and logerr arguments passed into each setup.py file.

Note: If you're looking for logs from our official benchmark rounds, see Round 9 and Round 8

Contribution Guidelines

The community has consistently helped in making these tests better, and we welcome any and all changes. These guidelines prevent us from having to give repeated feedback on the same topics:

• Use specific versions: If you're updating any software or dependency, please be specific with the version number. Also, update the appropriate README to reflect that change. Don't rely on the package manager to deliver a specific version, apt consistently returns different versions on Ubuntu 12.04 vs 14.04.
• Rope in experts: If you're making a performance tweak, our team may not be able to verify your code--we are not experts in every language. It's always helpful to ping expert users and provide a basic introduction on their credentials. If you are an expert that is willing to be pinged occasionally, please add yourself to the appropriate test README files.
• Use a personal Travis-CI account: This one is mainly for your own sanity. Our main Travis-CI can occasionally become clogged with so many pull requests that it takes a day to finish all the builds. If you create a fork and enable Travis-CI.org, you will get your own build queue. This means 1) only your commits/branches are being verified, so there is no delay waiting for an unrelated pull request, and 2) you can cancel unneeded items. This does not affect our own Travis-CI setup at all - any commits added to a pull request will be verifed as normal.
• Read the README: We know that's cliche. However, our toolset drags in a lot of different concepts and frameworks, and it can really help to read the README's, such as this one, the one inside the toolset/ directory, and the ones inside specific framework directories
• Use the Development Virtual Machine: Our Vagrant scripts can setup a VM for you that looks nearly identical to our test environment. This is even better than relying on the Travis-CI verification, and you are strongly encouraged to use this. See the deployment directory for specifics

---

Adding Frameworks or Tests

When adding a new framework or new test to an existing framework, please follow these steps:

• Update/add a benchmark_config
• Update/add a setup file
• Update/add an install.sh file
• (Optional) Update/add a bash_profile.sh file
• When creating a database test, use the table/collection hello_world. Our database setup scripts are stored inside the config/ folder if you need to see the database schema

The benchmark_config File

The benchmark_config file is used by our scripts to identify available tests - it should exist at the root of the framework directory. We

Here is an example benchmark_config from the Compojure framework. There are two different tests listed for the Compojure framework,

{
  "framework": "compojure",
  "tests": [{
    "default": {
      "setup_file": "setup",
      "json_url": "/compojure/json",
      "db_url": "/compojure/db/1",
      "query_url": "/compojure/db/",
      "fortune_url": "/compojure/fortune-hiccup",
      "plaintext_url": "/compojure/plaintext",
      "port": 8080,
      "approach": "Realistic",
      "classification": "Micro",
      "database": "MySQL",
      "framework": "compojure",
      "language": "Clojure",
      "orm": "Micro",
      "platform": "Servlet",
      "webserver": "Resin",
      "os": "Linux",
      "database_os": "Linux",
      "display_name": "compojure",
      "notes": "",
      "versus": "servlet"
    },
    "raw": {
      "setup_file": "setup",
      "db_url": "/compojure/dbraw/1",
      "query_url": "/compojure/dbraw/",
      "port": 8080,
      "approach": "Realistic",
      "classification": "Micro",
      "database": "MySQL",
      "framework": "compojure",
      "language": "Clojure",
      "orm": "Raw",
      "platform": "Servlet",
      "webserver": "Resin",
      "os": "Linux",
      "database_os": "Linux",
      "display_name": "compojure-raw",
      "notes": "",
      "versus": "servlet"
    }
  }]
}

• framework: Specifies the framework name.
• tests: A list of tests that can be run for this framework. In many cases, this contains a single element for the "default" test, but additional tests can be specified. Each test name must be unique when concatenated with the framework name. Each test will be run separately in our Rounds, so it is to your benefit to provide multiple variations in case one works better in some cases
  • setup_file: The location of the python setup file that can start and stop the test, excluding the .py ending. If your different tests require different setup approachs, use another setup file.
  • json_url (optional): The URI to the JSON test, typically /json
  • db_url (optional): The URI to the database test, typically /db
  • query_url (optional): The URI to the variable query test. The URI must be set up so that an integer can be applied to the end of the URI to specify the number of queries to run. For example, /query?queries= (to yield /query?queries=20) or /query/ (to yield /query/20)
  • fortune_url (optional): the URI to the fortunes test, typically /fortune
  • update_url (optional): the URI to the updates test, setup in a manner similar to query_url described above.
  • plaintext_url (optional): the URI of the plaintext test, typically /plaintext
  • port: The port the server is listening on
  • approach (metadata): Realistic or Stripped (see here for a description of all metadata attributes)
  • classification (metadata): Full, Micro, or Platform
  • database (metadata): MySQL, Postgres, MongoDB, SQLServer, or None
  • framework (metadata): name of the framework
  • language (metadata): name of the language
  • orm (metadata): Full, Micro, or Raw
  • platform (metadata): name of the platform
  • webserver (metadata): name of the web-server (also referred to as the "front-end server")
  • os (metadata): The application server's operating system, Linux or Windows
  • database_os (metadata): The database server's operating system, Linux or Windows
  • display_name (metadata): How to render this test permutation's name on the results web site. Some permutation names can be really long, so the display_name is provided in order to provide something more succinct.
  • versus (optional): The name of another test (elsewhere in this project) that is a subset of this framework. This allows for the generation of the framework efficiency chart in the results web site. For example, Compojure is compared to "servlet" since Compojure is built on the Servlets platform.

The requirements section here explains the expected response for each URL as well all metadata options available.

Testing on both Windows and Linux

If your framework and platform can execute on both Windows and Linux, we encourage you to specify tests for both operating systems. This increases the amount of testing you should do before submitting your pull-request, however, so we understand if you start with just one of the two. Travis-CI cannot automatically verify Windows-based tests, and therefore you should verify your code manually.

The steps involved are:

• Assuming you have implemented the Linux test already, add a new test permutation to your benchmark_config file for the Windows test. When the benchmark script runs on Linux, it skips tests where os in Windows and vice versa.
• Add the necessary tweaks to your setup file to start and stop on the new operating system. See, for example, the script for Go.
• Test on Windows and Linux to make sure everything works as expected.

Install File

The install.sh file for each framework starts the bash process which will install that framework. Typically, the first thing done is to call fw_depends to run installations for any necessary software that TFB has already created installation scripts for. TFB provides a reasonably wide range of core software, so your install.sh may only need to call fw_depends and exit. Note: fw_depends does not guarantee dependency installation, so list software in the proper order e.g. if foo depends on bar use fw_depends bar foo.

Here are some example install.sh files

#!/bin/bash

# My framework only needs nodejs
fw_depends nodejs

#!/bin/bash

# My framework needs nodejs and mono and go
fw_depends nodejs mono go

#!/bin/bash

# My framework needs nodejs
fw_depends nodejs

# ...and some other software that there is no installer script for.
# Note: Use IROOT variable to put software in the right folder. 
#       You can also use FWROOT to refer to the project root, or 
#       TROOT to refer to the root of your framework
# Please see guidelines on writing installation scripts
wget mystuff.tar.gz -O mystuff.tar.gz
untar mystuff.tar.gz
cd mystuff
make --prefix=$IROOT && sudo make install

To see what TFB provides installations for, look in toolset/setup/linux in the folders frameworks, languages, systools, and webservers. You should pass the filename, without the ".sh" extension, to fw_depends. Here is a listing as of July 2014:

$ ls frameworks                                                                
grails.sh  nawak.sh  play1.sh  siena.sh     vertx.sh  yesod.sh
jester.sh  onion.sh  play2.sh  treefrog.sh  wt.sh
$ ls languages
composer.sh  erlang.sh   hhvm.sh   mono.sh    perl.sh     pypy.sh     racket.sh   urweb.sh
dart.sh      go.sh       java.sh   nimrod.sh  python2.sh  ringojs.sh  xsp.sh
elixir.sh    haskell.sh  jruby.sh  nodejs.sh  php.sh      python3.sh  ruby.sh 
$ ls systools
leiningen.sh  maven.sh
$ ls webservers
lapis.sh  mongrel2.sh  nginx.sh  openresty.sh  resin.sh  weber.sh  zeromq.sh

Bash Environment File

The bash_profile.sh file is sourced before installing software or before running the framework test. This is mostly used when running your framework, to perform actions such as updating PATH or defining environment variables your framework requires e.g. GOROOT. You can use these variables:

• FWROOT: Root of project
• IROOT: Root of installation for the current framework
• TROOT: Root directory for the current framework

Example of bash_profile.sh:

# Set the root of our go installation
export GOROOT=${IROOT}/go

# Where to find the go executable
export PATH="$GOROOT/bin:$PATH"

export GOPATH=${FWROOT}/go

Do not cause any output, such as using echo, inside of bash_profile.sh

Setup Files

The setup file is responsible for starting and stopping the test. This script is responsible for (among other things):

• Modifying the framework's configuration to point to the correct database host
• Compiling and/or packaging the code (if impossible to do in install.sh)
• Starting the server
• Stopping the server

The setup file is a python script that contains a start() and a stop() function.
The start function should build the source, make any necessary changes to the framework's configuration, and then start the server. The stop function should shutdown the server, including all sub-processes as applicable.

Configuring database connectivity in start()

By convention, the configuration files used by a framework should specify the database server as localhost so that developing tests in a single-machine environment can be done in an ad hoc fashion, without using the benchmark scripts.

When running a benchmark script, the script needs to modify each framework's configuration so that the framework connects to a database host provided as a command line argument. In order to do this, use setup_util.replace_text() to make modifications prior to starting the server.

For example:

setup_util.replace_text("wicket/src/main/webapp/WEB-INF/resin-web.xml", "mysql:\/\/.*:3306", "mysql://" + args.database_host + ":3306")

Note: args contains a number of useful items, such as troot, iroot, fwroot (comparable to their bash counterparts in install.sh, database_host, client_host, and many others)

Note: Using localhost in the raw configuration file is not a requirement as long as the replace_text call properly injects the database host provided to the benchmark toolset as a command line argument.

A full example

Here is an example of Wicket's setup file.

import subprocess
import sys
import setup_util

##################################################
# start(args, logfile, errfile)
#
# Starts the server for Wicket
# returns 0 if everything completes, 1 otherwise
##################################################
def start(args, logfile, errfile):

# setting the database url
setup_util.replace_text(args.troot + "/src/main/webapp/WEB-INF/resin-web.xml", "mysql:\/\/.*:3306", "mysql://" + args.database_host + ":3306")

# 1. Compile and package
# 2. Clean out possible old tests
# 3. Copy package to Resin's webapp directory
# 4. Start resin
try:
  subprocess.check_call("mvn clean compile war:war", shell=True, cwd="wicket", stderr=errfile, stdout=logfile)
  subprocess.check_call("rm -rf $RESIN_HOME/webapps/*", shell=True, stderr=errfile, stdout=logfile)
  subprocess.check_call("cp $TROOT/target/hellowicket-1.0-SNAPSHOT.war $RESIN_HOME/webapps/wicket.war", shell=True, stderr=errfile, stdout=logfile)
  subprocess.check_call("$RESIN_HOME/bin/resinctl start", shell=True, stderr=errfile, stdout=logfile)
  return 0
except subprocess.CalledProcessError:
  return 1

##################################################
# stop(logfile, errfile)
#
# Stops the server for Wicket
# returns 0 if everything completes, 1 otherwise
##################################################
def stop(logfile):
try:
  subprocess.check_call("$RESIN_HOME/bin/resinctl shutdown", shell=True, stderr=errfile, stdout=logfile)
  return 0
except subprocess.CalledProcessError:
  return 1

A tool to generate your setup file

A contributor named @kpacha has built a pure JavaScript tool for generating the setup.py file for a new framework via an in-browser form. Check out his FrameworkBenchmarks Setup Builder.