Performance Testing Rails Applications ====================================== This guide covers the various ways of performance testing a Ruby on Rails application. After reading this guide, you will know: * The various types of benchmarking and profiling metrics. * How to generate performance and benchmarking tests. * How to install and use a GC-patched Ruby binary to measure memory usage and object allocation. * The benchmarking information provided by Rails inside the log files. * Various tools facilitating benchmarking and profiling. Performance testing is an integral part of the development cycle. It is very important that you don't make your end users wait for too long before the page is completely loaded. Ensuring a pleasant browsing experience for end users and cutting the cost of unnecessary hardware is important for any non-trivial web application. -------------------------------------------------------------------------------- Performance Test Cases ---------------------- Rails performance tests are a special type of integration tests, designed for benchmarking and profiling the test code. With performance tests, you can determine where your application's memory or speed problems are coming from, and get a more in-depth picture of those problems. In a freshly generated Rails application, `test/performance/browsing_test.rb` contains an example of a performance test: ```ruby require 'test_helper' require 'rails/performance_test_help' class BrowsingTest < ActionDispatch::PerformanceTest # Refer to the documentation for all available options # self.profile_options = { runs: 5, metrics: [:wall_time, :memory], # output: 'tmp/performance', formats: [:flat] } test "homepage" do get '/' end end ``` This example is a simple performance test case for profiling a GET request to the application's homepage. ### Generating Performance Tests Rails provides a generator called `performance_test` for creating new performance tests: ```bash $ rails generate performance_test homepage ``` This generates `homepage_test.rb` in the `test/performance` directory: ```ruby require 'test_helper' require 'rails/performance_test_help' class HomepageTest < ActionDispatch::PerformanceTest # Refer to the documentation for all available options # self.profile_options = { runs: 5, metrics: [:wall_time, :memory], # output: 'tmp/performance', formats: [:flat] } test "homepage" do get '/' end end ``` ### Examples Let's assume your application has the following controller and model: ```ruby # routes.rb root to: 'home#dashboard' resources :posts # home_controller.rb class HomeController < ApplicationController def dashboard @users = User.last_ten.includes(:avatars) @posts = Post.all_today end end # posts_controller.rb class PostsController < ApplicationController def create @post = Post.create(params[:post]) redirect_to(@post) end end # post.rb class Post < ActiveRecord::Base before_save :recalculate_costly_stats def slow_method # I fire gallzilion queries sleeping all around end private def recalculate_costly_stats # CPU heavy calculations end end ``` #### Controller Example Because performance tests are a special kind of integration test, you can use the `get` and `post` methods in them. Here's the performance test for `HomeController#dashboard` and `PostsController#create`: ```ruby require 'test_helper' require 'rails/performance_test_help' class PostPerformanceTest < ActionDispatch::PerformanceTest def setup # Application requires logged-in user login_as(:lifo) end test "homepage" do get '/dashboard' end test "creating new post" do post '/posts', post: { body: 'lifo is fooling you' } end end ``` You can find more details about the `get` and `post` methods in the [Testing Rails Applications](testing.html) guide. #### Model Example Even though the performance tests are integration tests and hence closer to the request/response cycle by nature, you can still performance test pure model code. Performance test for `Post` model: ```ruby require 'test_helper' require 'rails/performance_test_help' class PostModelTest < ActionDispatch::PerformanceTest test "creation" do Post.create body: 'still fooling you', cost: '100' end test "slow method" do # Using posts(:awesome) fixture posts(:awesome).slow_method end end ``` ### Modes Performance tests can be run in two modes: Benchmarking and Profiling. #### Benchmarking Benchmarking makes it easy to quickly gather a few metrics about each test run. By default, each test case is run **4 times** in benchmarking mode. To run performance tests in benchmarking mode: ```bash $ rake test:benchmark ``` #### Profiling Profiling allows you to make an in-depth analysis of each of your tests by using an external profiler. Depending on your Ruby interpreter, this profiler can be native (Rubinius, JRuby) or not (MRI, which uses RubyProf). By default, each test case is run **once** in profiling mode. To run performance tests in profiling mode: ```bash $ rake test:profile ``` ### Metrics Benchmarking and profiling run performance tests and give you multiple metrics. The availability of each metric is determined by the interpreter being used—none of them support all metrics—and by the mode in use. A brief description of each metric and their availability across interpreters/modes is given below. #### Wall Time Wall time measures the real world time elapsed during the test run. It is affected by any other processes concurrently running on the system. #### Process Time Process time measures the time taken by the process. It is unaffected by any other processes running concurrently on the same system. Hence, process time is likely to be constant for any given performance test, irrespective of the machine load. #### CPU Time Similar to process time, but leverages the more accurate CPU clock counter available on the Pentium and PowerPC platforms. #### User Time User time measures the amount of time the CPU spent in user-mode, i.e. within the process. This is not affected by other processes and by the time it possibly spends blocked. #### Memory Memory measures the amount of memory used for the performance test case. #### Objects Objects measures the number of objects allocated for the performance test case. #### GC Runs GC Runs measures the number of times GC was invoked for the performance test case. #### GC Time GC Time measures the amount of time spent in GC for the performance test case. #### Metric Availability ##### Benchmarking | Interpreter | Wall Time | Process Time | CPU Time | User Time | Memory | Objects | GC Runs | GC Time | | ------------ | --------- | ------------ | -------- | --------- | ------ | ------- | ------- | ------- | | **MRI** | yes | yes | yes | no | yes | yes | yes | yes | | **REE** | yes | yes | yes | no | yes | yes | yes | yes | | **Rubinius** | yes | no | no | no | yes | yes | yes | yes | | **JRuby** | yes | no | no | yes | yes | yes | yes | yes | ##### Profiling | Interpreter | Wall Time | Process Time | CPU Time | User Time | Memory | Objects | GC Runs | GC Time | | ------------ | --------- | ------------ | -------- | --------- | ------ | ------- | ------- | ------- | | **MRI** | yes | yes | no | no | yes | yes | yes | yes | | **REE** | yes | yes | no | no | yes | yes | yes | yes | | **Rubinius** | yes | no | no | no | no | no | no | no | | **JRuby** | yes | no | no | no | no | no | no | no | NOTE: To profile under JRuby you'll need to run `export JRUBY_OPTS="-Xlaunch.inproc=false --profile.api"` **before** the performance tests. ### Understanding the Output Performance tests generate different outputs inside `tmp/performance` directory depending on their mode and metric. #### Benchmarking In benchmarking mode, performance tests generate two types of outputs. ##### Command Line This is the primary form of output in benchmarking mode. Example: ```bash BrowsingTest#test_homepage (31 ms warmup) wall_time: 6 ms memory: 437.27 KB objects: 5,514 gc_runs: 0 gc_time: 19 ms ``` ##### CSV Files Performance test results are also appended to `.csv` files inside `tmp/performance`. For example, running the default `BrowsingTest#test_homepage` will generate following five files: * BrowsingTest#test_homepage_gc_runs.csv * BrowsingTest#test_homepage_gc_time.csv * BrowsingTest#test_homepage_memory.csv * BrowsingTest#test_homepage_objects.csv * BrowsingTest#test_homepage_wall_time.csv As the results are appended to these files each time the performance tests are run in benchmarking mode, you can collect data over a period of time. This can be very helpful in analyzing the effects of code changes. Sample output of `BrowsingTest#test_homepage_wall_time.csv`: ```bash measurement,created_at,app,rails,ruby,platform 0.00738224999999992,2009-01-08T03:40:29Z,,3.0.0,ruby-1.8.7.249,x86_64-linux 0.00755874999999984,2009-01-08T03:46:18Z,,3.0.0,ruby-1.8.7.249,x86_64-linux 0.00762099999999993,2009-01-08T03:49:25Z,,3.0.0,ruby-1.8.7.249,x86_64-linux 0.00603075000000008,2009-01-08T04:03:29Z,,3.0.0,ruby-1.8.7.249,x86_64-linux 0.00619899999999995,2009-01-08T04:03:53Z,,3.0.0,ruby-1.8.7.249,x86_64-linux 0.00755449999999991,2009-01-08T04:04:55Z,,3.0.0,ruby-1.8.7.249,x86_64-linux 0.00595999999999997,2009-01-08T04:05:06Z,,3.0.0,ruby-1.8.7.249,x86_64-linux 0.00740450000000004,2009-01-09T03:54:47Z,,3.0.0,ruby-1.8.7.249,x86_64-linux 0.00603150000000008,2009-01-09T03:54:57Z,,3.0.0,ruby-1.8.7.249,x86_64-linux 0.00771250000000012,2009-01-09T15:46:03Z,,3.0.0,ruby-1.8.7.249,x86_64-linux ``` #### Profiling In profiling mode, performance tests can generate multiple types of outputs. The command line output is always presented but support for the others is dependent on the interpreter in use. A brief description of each type and their availability across interpreters is given below. ##### Command Line This is a very basic form of output in profiling mode: ```bash BrowsingTest#test_homepage (58 ms warmup) process_time: 63 ms memory: 832.13 KB objects: 7,882 ``` ##### Flat Flat output shows the metric—time, memory, etc—measure in each method. [Check Ruby-Prof documentation for a better explanation](http://ruby-prof.rubyforge.org/files/examples/flat_txt.html). ##### Graph Graph output shows the metric measure in each method, which methods call it and which methods it calls. [Check Ruby-Prof documentation for a better explanation](http://ruby-prof.rubyforge.org/files/examples/graph_txt.html). ##### Tree Tree output is profiling information in calltree format for use by [kcachegrind](http://kcachegrind.sourceforge.net/html/Home.html) and similar tools. ##### Output Availability | | Flat | Graph | Tree | | ------------ | ---- | ----- | ---- | | **MRI** | yes | yes | yes | | **REE** | yes | yes | yes | | **Rubinius** | yes | yes | no | | **JRuby** | yes | yes | no | ### Tuning Test Runs Test runs can be tuned by setting the `profile_options` class variable on your test class. ```ruby require 'test_helper' require 'rails/performance_test_help' class BrowsingTest < ActionDispatch::PerformanceTest self.profile_options = { runs: 5, metrics: [:wall_time, :memory] } test "homepage" get '/' end end ``` In this example, the test would run 5 times and measure wall time and memory. There are a few configurable options: | Option | Description | Default | Mode | | ---------- | ------------------------------------------ | ----------------------------- | --------- | | `:runs` | Number of runs. | Benchmarking: 4, Profiling: 1 | Both | | `:output` | Directory to use when writing the results. | `tmp/performance` | Both | | `:metrics` | Metrics to use. | See below. | Both | | `:formats` | Formats to output to. | See below. | Profiling | Metrics and formats have different defaults depending on the interpreter in use. | Interpreter | Mode | Default metrics | Default formats | | -------------- | ------------ | ------------------------------------------------------- | ----------------------------------------------- | | **MRI/REE** | Benchmarking | `[:wall_time, :memory, :objects, :gc_runs, :gc_time]` | N/A | | | Profiling | `[:process_time, :memory, :objects]` | `[:flat, :graph_html, :call_tree, :call_stack]` | | **Rubinius** | Benchmarking | `[:wall_time, :memory, :objects, :gc_runs, :gc_time]` | N/A | | | Profiling | `[:wall_time]` | `[:flat, :graph]` | | **JRuby** | Benchmarking | `[:wall_time, :user_time, :memory, :gc_runs, :gc_time]` | N/A | | | Profiling | `[:wall_time]` | `[:flat, :graph]` | As you've probably noticed by now, metrics and formats are specified using a symbol array with each name [underscored.](http://api.rubyonrails.org/classes/String.html#method-i-underscore) ### Performance Test Environment Performance tests are run in the `test` environment. But running performance tests will set the following configuration parameters: ```bash ActionController::Base.perform_caching = true ActiveSupport::Dependencies.mechanism = :require Rails.logger.level = ActiveSupport::BufferedLogger::INFO ``` As `ActionController::Base.perform_caching` is set to `true`, performance tests will behave much as they do in the `production` environment. ### Installing GC-Patched MRI To get the best from Rails' performance tests under MRI, you'll need to build a special Ruby binary with some super powers. The recommended patches for each MRI version are: | Version | Patch | | --------------- | --------- | | 1.8.6 | ruby186gc | | 1.8.7 | ruby187gc | | 1.9.2 and above | gcdata | All of these can be found on [RVM's _patches_ directory](https://github.com/wayneeseguin/rvm/tree/master/patches/ruby) under each specific interpreter version. Concerning the installation itself, you can either do this easily by using [RVM](http://rvm.beginrescueend.com) or you can build everything from source, which is a little bit harder. #### Install Using RVM The process of installing a patched Ruby interpreter is very easy if you let RVM do the hard work. All of the following RVM commands will provide you with a patched Ruby interpreter: ```bash $ rvm install 1.9.2-p180 --patch gcdata $ rvm install 1.8.7 --patch ruby187gc $ rvm install 1.9.2-p180 --patch ~/Downloads/downloaded_gcdata_patch.patch ``` You can even keep your regular interpreter by assigning a name to the patched one: ```bash $ rvm install 1.9.2-p180 --patch gcdata --name gcdata $ rvm use 1.9.2-p180 # your regular ruby $ rvm use 1.9.2-p180-gcdata # your patched ruby ``` And it's done! You have installed a patched Ruby interpreter. #### Install From Source This process is a bit more complicated, but straightforward nonetheless. If you've never compiled a Ruby binary before, follow these steps to build a Ruby binary inside your home directory. ##### Download and Extract ```bash $ mkdir rubygc $ wget $ tar -xzvf $ cd ``` ##### Apply the Patch ```bash $ curl http://github.com/wayneeseguin/rvm/raw/master/patches/ruby/1.9.2/p180/gcdata.patch | patch -p0 # if you're on 1.9.2! $ curl http://github.com/wayneeseguin/rvm/raw/master/patches/ruby/1.8.7/ruby187gc.patch | patch -p0 # if you're on 1.8.7! ``` ##### Configure and Install The following will install Ruby in your home directory's `/rubygc` directory. Make sure to replace `` with a full patch to your actual home directory. ```bash $ ./configure --prefix=//rubygc $ make && make install ``` ##### Prepare Aliases For convenience, add the following lines in your `~/.profile`: ```bash alias gcruby='~/rubygc/bin/ruby' alias gcrake='~/rubygc/bin/rake' alias gcgem='~/rubygc/bin/gem' alias gcirb='~/rubygc/bin/irb' alias gcrails='~/rubygc/bin/rails' ``` Don't forget to use your aliases from now on. ### Using Ruby-Prof on MRI and REE Add Ruby-Prof to your applications' Gemfile if you want to benchmark/profile under MRI or REE: ```ruby gem 'ruby-prof' ``` Now run `bundle install` and you're ready to go. Command Line Tools ------------------ Writing performance test cases could be an overkill when you are looking for one time tests. Rails ships with two command line tools that enable quick and dirty performance testing: ### `benchmarker` Usage: ```bash Usage: rails benchmarker 'Ruby.code' 'Ruby.more_code' ... [OPTS] -r, --runs N Number of runs. Default: 4 -o, --output PATH Directory to use when writing the results. Default: tmp/performance -m, --metrics a,b,c Metrics to use. Default: wall_time,memory,objects,gc_runs,gc_time ``` Example: ```bash $ rails benchmarker 'Item.all' 'CouchItem.all' --runs 3 --metrics wall_time,memory ``` ### `profiler` Usage: ```bash Usage: rails profiler 'Ruby.code' 'Ruby.more_code' ... [OPTS] -r, --runs N Number of runs. Default: 1 -o, --output PATH Directory to use when writing the results. Default: tmp/performance -m, --metrics a,b,c Metrics to use. Default: process_time,memory,objects -f, --formats x,y,z Formats to output to. Default: flat,graph_html,call_tree ``` Example: ```bash $ rails profiler 'Item.all' 'CouchItem.all' --runs 2 --metrics process_time --formats flat ``` NOTE: Metrics and formats vary from interpreter to interpreter. Pass `--help` to each tool to see the defaults for your interpreter. Helper Methods -------------- Rails provides various helper methods inside Active Record, Action Controller and Action View to measure the time taken by a given piece of code. The method is called `benchmark()` in all the three components. ### Model ```ruby Project.benchmark("Creating project") do project = Project.create("name" => "stuff") project.create_manager("name" => "David") project.milestones << Milestone.all end ``` This benchmarks the code enclosed in the `Project.benchmark("Creating project") do...end` block and prints the result to the log file: ```ruby Creating project (185.3ms) ``` Please refer to the [API docs](http://api.rubyonrails.org/classes/ActiveSupport/Benchmarkable.html#method-i-benchmark) for additional options to `benchmark()`. ### Controller Similarly, you could use this helper method inside [controllers.](http://api.rubyonrails.org/classes/ActiveSupport/Benchmarkable.html) ```ruby def process_projects benchmark("Processing projects") do Project.process(params[:project_ids]) Project.update_cached_projects end end ``` NOTE: `benchmark` is a class method inside controllers. ### View And in [views](http://api.rubyonrails.org/classes/ActiveSupport/Benchmarkable.html:) ```erb <% benchmark("Showing projects partial") do %> <%= render @projects %> <% end %> ``` Request Logging --------------- Rails log files contain very useful information about the time taken to serve each request. Here's a typical log file entry: ```bash Processing ItemsController#index (for 127.0.0.1 at 2009-01-08 03:06:39) [GET] Rendering template within layouts/items Rendering items/index Completed in 5ms (View: 2, DB: 0) | 200 OK [http://0.0.0.0/items] ``` For this section, we're only interested in the last line: ```bash Completed in 5ms (View: 2, DB: 0) | 200 OK [http://0.0.0.0/items] ``` This data is fairly straightforward to understand. Rails uses millisecond(ms) as the metric to measure the time taken. The complete request spent 5 ms inside Rails, out of which 2 ms were spent rendering views and none was spent communication with the database. It's safe to assume that the remaining 3 ms were spent inside the controller. Michael Koziarski has an [interesting blog post](http://www.therailsway.com/2009/1/6/requests-per-second) explaining the importance of using milliseconds as the metric. Useful Links ------------ ### Rails Plugins and Gems * [Rails Analyzer](http://rails-analyzer.rubyforge.org) * [Rails Footnotes](https://github.com/josevalim/rails-footnotes/tree/master) * [Query Reviewer](https://github.com/nesquena/query_reviewer) * [MiniProfiler](http://www.miniprofiler.com) ### Generic Tools * [httperf](http://www.hpl.hp.com/research/linux/httperf/) * [ab](http://httpd.apache.org/docs/2.2/programs/ab.html) * [JMeter](http://jakarta.apache.org/jmeter/) * [kcachegrind](http://kcachegrind.sourceforge.net/html/Home.html) ### Tutorials and Documentation * [ruby-prof API Documentation](http://ruby-prof.rubyforge.org) * [Request Profiling Railscast](http://railscasts.com/episodes/98-request-profiling) - Outdated, but useful for understanding call graphs. Commercial Products ------------------- Rails has been lucky to have a few companies dedicated to Rails-specific performance tools. A couple of those are: * [New Relic](http://www.newrelic.com) * [Scout](http://scoutapp.com)