Ruby 2.1

Tracing Object Allocations

Sam Rawlins / @srawlins

Ruby 2.1 is out!

• $ rbenv install 2.1
• $ rvm install ruby-2.1

Ruby 2.1 news

= NEWS for Ruby 2.1.0

This document is a list of user visible feature changes made between
releases except for bug fixes.

Note that each entry is kept so brief that no reason behind or
reference information is supplied with.  For a full list of changes
with all sufficient information, see the ChangeLog file.

== Changes since the 2.0.0 release

=== Language changes

* Now the default values of keyword arguments can be omitted.  Those
  "required keyword arguments" need giving explicitly at the call time.

* Added suffixes for integer and float literals: 'r', 'i', and 'ri'.
  * "42r" and "3.14r" are evaluated as Rational(42, 1) and 3.14.rationalize,
    respectively.  But exponential form with 'r' suffix like "6.022e+23r" is
    not accepted because it is misleading.
  * "42i" and "3.14i" are evaluated as Complex(0, 42) and Complex(0, 3.14),
    respectively.
  * "42ri" and "3.14ri" are evaluated as Complex(0, 42r) and Complex(0, 3.14r),
    respectively.

* def-expr now returns the symbol of its name instead of nil.

* Added 'f' suffix for string literals that returns a frozen String object.

=== Core classes updates (outstanding ones only)

* Binding
  * New methods
    * Binding#local_variable_get(symbol)
    * Binding#local_variable_set(symbol, obj)
    * Binding#local_variable_defined?(symbol)

* GC
  * added environment variable:
    * RUBY_HEAP_SLOTS_GROWTH_FACTOR: growth rate of the heap.

* Integer
  * New methods
    * Fixnum#bit_length
    * Bignum#bit_length
  * Bignum performance improvement
    * Use GMP if available.
      GMP is used only for several operations:
      multiplication, division, radix conversion, GCD

* IO
  * extended methods:
    * IO#seek supports SEEK_DATA and SEEK_HOLE as whence.
    * IO#seek accepts symbols (:CUR, :END, :SET, :DATA, :HOLE) for 2nd argument.
    * IO#read_nonblock accepts optional `exception: false` to return symbols
    * IO#write_nonblock accepts optional `exception: false` to return symbols

* Kernel
  * New methods:
    * Kernel#singleton_method

* Module
  * New methods:
    * Module#using, which activates refinements of the specified module only
      in the current class or module definition.
    * Module#singleton_class? returns true if the receiver is a singleton class
      or false if it is an ordinary class or module.
  * extended methods:
    * Module#refine is no longer experimental.
    * Module#include and Module#prepend are now public methods.

* Mutex
  * misc
    * Mutex#owned? is no longer experimental.

* Numeric
  * extended methods:
    * Numeric#step allows the limit argument to be omitted, in which
      case an infinite sequence of numbers is generated.  Keyword
      arguments `to` and `by` are introduced for ease of use.

* Process
  * New methods:
    * alternative methods to $0/$0=:
      * Process.argv0() returns the original value of $0.
      * Process.setproctitle() sets the process title without affecting $0.
    * Process.clock_gettime
    * Process.clock_getres

* RbConfig
  * New constants:
    * RbConfig::SIZEOF is added to provide the size of C types.

* String
  * New methods:
    * String#scrub and String#scrub! verify and fix invalid byte sequence.
  * extended methods:
    * If invalid: :replace is specified for String#encode, replace
      invalid byte sequence even if the destination encoding equals to
      the source encoding.

* Symbol
  * All symbols are now frozen.

* pack/unpack (Array/String)
  * Q! and q! directives for long long type if platform has the type.

* toplevel
  * extended methods:
    * main.using is no longer experimental. The method activates refinements
      in the ancestors of the argument module to support refinement
      inheritance by Module#include.

=== Core classes compatibility issues (excluding feature bug fixes)

* IO
  * incompatible changes:
    * open ignore internal encoding if external encoding is ASCII-8BIT.

* Kernel#eval, Kernel#instance_eval, and Module#module_eval.
  * Copies the scope information of the original environment, which means
    that private, protected, public, and module_function without arguments
    do not affect the environment outside the eval string.
    For example, `class Foo; eval "private"; def foo; end; end' doesn't make
    Foo#foo private.

* Kernel#untrusted?, untrust, and trust
  * These methods are deprecated and their behavior is same as tainted?,
    taint, and untaint, respectively.  If $VERBOSE is true, they show warnings.

* Module#ancestors
  * The ancestors of a singleton class now include singleton classes,
    in particular itself.

* Module#define_method and Object#define_singleton_method
  * Now they return the symbols of the defined methods, not the methods/procs
    themselves.

* Numeric#quo
  * Raises TypeError instead of ArgumentError if the receiver doesn't have
    to_r method.

* Proc
  * Returning from lambda proc now always exits from the Proc, not from the
    method where the lambda is created.  Returing from non-lambda proc exits
    from the method, same as the former behavior.


=== Stdlib updates (outstanding ones only)

* CGI::Util
  * All class methods modulized.

* Digest
  * extended methods:
    * Digest::Class.file takes optional arguments for its constructor

* Matrix
  * Added Vector#cross_product.

* Net::SMTP
  * Added Net::SMTP#rset to implement the RSET command

* Pathname
  * New methods:
    * Pathname#write
    * Pathname#binwrite

* OpenSSL::BN
  * extended methods:
    * OpenSSL::BN.new allows Fixnum/Bignum argument.

* open-uri
  * Support multiple fields with same field name (like Set-Cookie).

* RDoc
  * Updated to 4.1.0.preview.1.  Major enhancements include a modified default
    template and accessibility enhancements.

    For a list of minor enhancements and bug fixes see:
    https://github.com/rdoc/rdoc/blob/v4.1.0.preview.1/History.rdoc

* Resolv
  * New methods:
    * Resolv::DNS.fetch_resource
  * One-shot multicast DNS support
  * Support LOC resources

* REXML::Parsers::SAX2Parser
  * Fixes wrong number of arguments of entitydecl event. Document of the event
    says "an array of the entity declaration" but implemenation passes two
    or more arguments. It is an implementation bug but it breaks backword
    compatibility.

* REXML::Parsers::StreamParser
  * Supports "entity" event.

* REXML::Text
  * REXML::Text#<< supports method chain like 'text << "XXX" << "YYY"'.
  * REXML::Text#<< supports not "raw" mode.

* Rinda::RingServer, Rinda::RingFinger
  * Rinda now supports multicast sockets.  See Rinda::RingServer and
    Rinda::RingFinger for details.

* RubyGems
  * Updated to 2.2.0.preview.1  For a list of enhancements and bug fixes see:
    https://github.com/rubygems/rubygems/blob/v2.2.0.preview.1/History.txt

* Set
  * New methods:
    * Set#intersect?
    * Set#disjoint?

* Socket
  * New methods:
    * Socket.getifaddrs

* StringScanner
  * extended methods:
    * StringScanner#[] supports named captures.

* Syslog::Logger
  * Added facility.

* Tempfile
  * New methods:
    * Tempfile.create


* WEBrick
  * The body of a response may now be a StringIO or other IO-like that responds
    to #readpartial and #read.

* XMLRPC::Client
  * New methods:
    * XMLRPC::Client#http. It returns Net::HTTP for the client. Normally,
      it is not needed. It is useful when you want to change minor HTTP client
      options. You can change major HTTP client options by XMLRPC::Client
      methods. You should use XMLRPC::Client methods for changing major
      HTTP client options instead of XMLRPC::Client#http.


=== Stdlib compatibility issues (excluding feature bug fixes)

* objspace
  * new method:
    * ObjectSpace.trace_object_allocations
    * ObjectSpace.trace_object_allocations_start
    * ObjectSpace.trace_object_allocations_stop
    * ObjectSpace.trace_object_allocations_clear
    * ObjectSpace.allocation_sourcefile
    * ObjectSpace.allocation_sourceline
    * ObjectSpace.allocation_class_path
    * ObjectSpace.allocation_method_id
    * ObjectSpace.allocation_generation

* Set
  * incompatible changes:
    * Set#to_set now returns self instead of generating a copy.

* URI
  * incompatible changes:
    * URI.decode_www_form follows current WHATWG URL Standard.
      It gets encoding argument to specify the character encoding.
      It now allows loose percent encoded strings, but denies ;-separator.
    * URI.encode_www_form follows current WHATWG URL Standard.
      It gets encoding argument to convert before percent encode.
      UTF-16 strings aren't converted to UTF-8 before percent encode by default.


=== Built-in global variables compatibility issues

* $SAFE
  * $SAFE=4 is obsolete.  If $SAFE is set to 4 or larger, an ArgumentError
    is raised.
  

* objspace
  * new method:
    * ObjectSpace.trace_object_allocations
    * ObjectSpace.trace_object_allocations_start
    * ObjectSpace.trace_object_allocations_stop
    * ObjectSpace.trace_object_allocations_clear
    * ObjectSpace.allocation_sourcefile
    * ObjectSpace.allocation_sourceline
    * ObjectSpace.allocation_class_path
    * ObjectSpace.allocation_method_id
    * ObjectSpace.allocation_generation
    

ObjectSpace is not new

Mostly known for

• ObjectSpace.count_objects
• ObjectSpace.each_object
• ObjectSpace.garbage_collect

ObjectSpace.new_in_ruby_2_1

• ObjectSpace.trace_object_allocations
• ObjectSpace.trace_object_allocations_start
• ObjectSpace.trace_object_allocations_stop
• ObjectSpace.trace_object_allocations_clear
• ...

Anecdote from GitHub

GitHub.preload_all
GC.start
count = ObjectSpace.count_objects

puts count[:TOTAL] - count[:FREE]
#=> 605183

> 600k Ruby objects!

"Where am I supposedly hogging all this memory?"

Enter ::trace_object_allocations

# example.rb

1  class MyClass
2    def an_array
3      return [2,3,5,7,11,13,17,19,23,29,31]
4    end

5    def a_string
6      return "a String"
7    end
8  end

require 'objspace'

a = s = nil

require 'objspace'

Why?

• Reduce memory footprint
• Reduce garbage collection time
  • (marking and sweeping)

But my application isn't on Ruby 2.1!

That's okay! It's a diagnostic tool!

ObjectSpace.trace_object_allocations

is

Limited Fine-Grained Just the Start

The Next Step:

Aggregation

AllocationStats

requires Ruby 2.1

Low-level Example

 1  class MyClass
 2    def my_method

 7
 8  require "allocation_stats"
 9  stats = AllocationStats.trace do

12
13  puts stats.allocations.to_text

           sourcefile             sourceline  class_path  method_id   class
--------------------------------  ----------  ----------  ---------  -------

group_by Example

 1  class MyClass
 2    def my_method

 7
 8  require "allocation_stats"
 9  stats = AllocationStats.trace do
10    MyClass.new.my_method
11  end

12
13  puts stats.allocations

             sourcefile                sourceline   class   count
-------------------------------------  ----------  -------  -----
./examples/trace_my_class_group_by.rb           4  String       1
./examples/trace_my_class_group_by.rb           3  Hash         1

Psych

require "yaml"
require "allocation_stats"

stats = AllocationStats.trace do
  # lots of objects from Rbconfig::CONFIG["rubylibdir"]
  

               sourcefile                             class               count
----------------------------------------  ------------------------------  -----

Psych - sorted

require "yaml"
require "allocation_stats"

stats = AllocationStats.trace do
  # lots of objects from Rbconfig::CONFIG["rubylibdir"]
  y = YAML.dump(["one string", "two string"])
end

puts stats.allocations(alias_paths: true).
  

               sourcefile                             class               count
-----------------------------------------  ------------------------------  -----

class_plus

require 'hike'
require 'allocation_stats'

stats = AllocationStats.trace do

                   sourcefile                              class_plus           count
-------------------------------------------------  ---------------------------  -----

burn

require 'hike'
require 'allocation_stats'

stats = AllocationStats.new

                     sourcefile                          class_plus     count
-----------------------------------------------------  ---------------  -----
<RUBYLIBDIR>/pathname.rb                               String             287
<RUBYLIBDIR>/pathname.rb                               Array<String>       73
<GEMDIR>/gems/hike-1.2.3/lib/hike/index.rb             String              73
<RUBYLIBDIR>/pathname.rb                               MatchData           13
<GEMDIR>/gems/hike-1.2.3/lib/hike/index.rb             Pathname            12
<GEMDIR>/gems/hike-1.2.3/lib/hike/index.rb             Array<Pathname>     11
<GEMDIR>/gems/hike-1.2.3/lib/hike/paths.rb             String               7
<GEMDIR>/gems/hike-1.2.3/lib/hike/index.rb             Array<String>        6
<GEMDIR>/gems/hike-1.2.3/lib/hike/index.rb             MatchData            6
<GEMDIR>/gems/hike-1.2.3/lib/hike/trail.rb             String               6
<GEMDIR>/gems/hike-1.2.3/lib/hike/trail.rb             Array<String>        5
<GEMDIR>/gems/hike-1.2.3/lib/hike/index.rb             Hash                 5
...

Run inside your specs

# top of your spec_helper.rb

require 'allocation_stats'
AllocationStats.trace_rspec

mail$ be rspec spec/mail
...
Finished in 4 minutes 42.52 seconds
1469 examples, 0 failures, 11 pending
...
Top 10 allocation sites:
  String allocations at /lib/mail/utilities.rb:180
    4639 allocations during ./spec/mail/attachments_list_spec.rb:312
    4443 allocations during ./spec/mail/message_spec.rb:1627
    3975 allocations during ./spec/mail/network/retriever_methods/test_retriever_spec.rb:78
  Array allocations at /lib/mail/parsers/received_parser.rb:13
    2326 allocations during ./spec/mail/message_spec.rb:1627
    1690 allocations during ./spec/mail/example_emails_spec.rb:300
  ...
Total allocations: 855826

Object Churn

Expensive Garbage Collection

per-Rack::Request allocations

Rack AllocationStats

Rack Middleware

A Simple Sinatra App localhost:9292/erb localhost:9292/erb?ras[trace]=true  

class SinatraTemplatesApp < Sinatra::Base

?ras[help]

Rack AllocationStats help

Append ras[trace]=true to your URL to trace all object allocations during the
request. The server will respond with statistics on the allocations, instead of
the application's response.

Options

ras[times]=COUNT
    send the request to the application COUNT times, to reduce statistics from
    initial allocations that don't happen on repeat requests.

ras[help]
    immediately respond with this help text

ras[scope]=SCOPE
    limits the list of allocations in the response to those whose `sourcefile`
    matches SCOPE. `ras[scope]='.'` is a special case which is expanded to be
    the current working directory.

...

?ras[alias_paths]=true

?ras[times]=10

?ras[output]=json

What do I do???

Allocate less.

How do I allocate less?

That depends :)

Use self-modifying collection methods where possible.

diff --git a/lib/temple/hash.rb b/lib/temple/hash.rb
index a2235c0..f1e5a70 100644
--- a/lib/temple/hash.rb
+++ b/lib/temple/hash.rb
@@ -24,3 +24,3 @@ def each
     def keys
-      @hash.inject([]) {|keys, h| keys += h.keys }.uniq
+      @hash.inject([]) {|keys, h| keys.concat(h.keys) }.uniq
     end

Memoization

(Especially String-building)

String#freeze

(also Ruby 2.1)

String literals, when frozen, will always refer to the same object.

String#freeze

STRING = "This is a static string"
def no_freeze
  STRING.tr(" ", "_") << "not frozen"
end

def freeze
  STRING.tr(" ".freeze, "_".freeze) << "frozen".freeze
end

  user   system   total    real       gc    allocations
  9.380   0.100   9.480 (  9.502)    1.724     1600000
  7.110   0.050   7.160 (  7.190)    1.098      400000

Not Always Easy

action_view/helpers/tag_helper.rb:65

64 def tag(name, options = nil, open = false, escape = true)
65  "<#{name}#{tag_options(options, escape) if options}#{open ? ">" : "/>"}".html_safe
66 end

Not Always Easy

sqlite3/statement.rb:108

106 def each
107   loop do

• Where is #step?
• It's an 88-line C function in ext/sqlite3/statement.c :(

Be Excited!

• Be aware of your object allocations.
• Be aware of how much garbage collection costs you.
• Add ::trace_object_allocations to your toolbox.

Watch This Space!

Things to Google

• github blog hey judy
• ruby allocation_stats
• ruby rack-allocation_stats
• koichi sasada efficient 2.1

Thanks

• Aman Gupta / @tmm1
• GitHub
• Matt Brooks
• Mike Moore for MWRC
• Hakim El Hattab for reveal.js