MRI Memory Management

Among the most notable trends in the Ruby community nowadays there is an odd one that can be described as «1000 and 1 the most sophisticated and intricate ways to despise MRI’s garbage collector». Let’s figure out why there are so many people who hate it, how interpreter’s memory management works in general and what all this means to a down-to-earth Ruby developer who just needs to fix that goddamn memory leak.

Transcript

2. Memory management in MRI
3. About me Arthur Pirogovski arthur@flyintealeaf.com
4. About me Coding for , mostly Ruby.
5. Ruby community
6. Ruby community loves Ruby
7. Ruby community hates Ruby (at times)
8. What’s wrong with Ruby?
9. What’s wrong with Ruby? • not as fast as some languages
10. What’s wrong with Ruby? • not as fast as some languages • concurrency is a swear word
11. What’s wrong with Ruby? • not as fast as some languages • concurrency is a swear word • garbage collector is garbage
12. Garbage Collector is garbage Y?
13. Lifetime of an object
14. Lifetime of an object Object.new
15. Lifetime of an object Object.new rb_class_new_instance()
16. Lifetime of an object Object.new rb_obj_alloc() 㱺 .allocate rb_obj_call_init() 㱺 #initialize
17. Lifetime of an object .allocate rb_newobj() /* gc.c */
18. Lifetime of an object rb_newobj(): 1. try to bite from freelist
19. Lifetime of an object rb_newobj(): 1. try to bite from freelist 2. if freelist is empty invoke GC
20. Lifetime of an object rb_newobj(): 1. try to bite from freelist (except for basic types) 2. if freelist is empty invoke GC
21. Lifetime of an object rb_newobj(): 1. try to bite from freelist (except for basic types) 2. if freelist is empty invoke GC (everything lives on a heap)
22. Native Types • Custom allocators • Embedded values (boxing)
23. Fixnum A 31/63-bit number where the original number is lshifted one bit and LSB is set to “1”
24. true, false, nil C constants: 2, 0, 4
25. String embedded if length < 23 chars (~ 11-12 chars for 32-bit) CoW for #dup
26. Symbol is stored in the symbol table is not GCed
27. 1. Stop the World GC in 1.9
28. 1. Stop the World 2. Mark all the linked objects GC in 1.9
29. 1. Stop the World 2. Mark all the linked objects 3. Sweep until enough GC in 1.9
30. 1. Stop the World 2. Mark all the linked objects 3. Sweep until enough 4. malloc() if not enough GC in 1.9
31. 1. Stop the World 2. Mark all the linked objects 3. Sweep until enough 4. malloc() if not enough 5. Wake up the World GC in 1.9
32. • Blocks all threads GC problems
33. • Blocks all threads • Is prone to ghost references GC problems
34. • Blocks all threads • Is prone to ghost references • Is slow GC problems
35. the less often it runs, the slower each run is GC slowness
36. the more you parallelize, the more GC impacts you GC slowness
37. the more you parallelize, the more GC impacts you (think EventMachine et al) GC slowness
38. How can we minimize GC impact?
39. How can we minimize GC impact? - Allocate less
40. Hash[something.map] allocates: 1. An array 2. A hash Allocate less
41. something.reduce({}) allocates: 1. An array 2. A hash Allocate less
42. avoid loading unnecessary libs which lines in Gemfile can be removed? Allocate less
43. know your allocations what’s hidden in the memory of your Ruby app? Allocate less
44. Sometimes it’s really hard to figure that out
45. In Ruby we have:
46. • no working memory profiler In Ruby we have:
47. • no working memory profiler • no memory profiling API In Ruby we have:
48. • no working memory profiler • no memory profiling API • borked finalizers In Ruby we have:
49. • no working memory profiler • no memory profiling API • borked finalizers • ObjectSpace! In Ruby we have:
50. ObjectSpace.count_objects {:TOTAL=>101383, :FREE=>164, :T_OBJECT=>6822, :T_CLASS=>865, :T_FLOAT=>7, :T_STRING=>63301, ...} helps to find leaks in native Ruby types like String ObjectSpace?
51. ObjectSpace.each_object. reduce(Hash.new(0)) { |hash, o| hash[o.class] +=1; hash } helps to find leaks in custom Ruby types like ActiveRecord ObjectSpace?
52. it works, but it’s: • slow • intrusive (allocates objects) • unable to show code location ObjectSpace?
53. Functions have .source_location: ActiveRecord::Base. instance_method(:reload). source_location ... seamless_database_pool_adapter.rb", 80] Showing code location
54. No such thing to reveal where an object was allocated ;( Showing code location
55. No such thing to reveal where an object was allocated But we can hack it up! Showing code location
56. class M attr_reader :allocated_at def initialize @allocated_at = caller.first end end m = M.new puts m.allocated_at.inspect Showing code location
57. class BasicObject attr_reader :allocated_at def initialize_with_allocated_at(*options) unless @allocated_at @allocated_at = caller.first return initialize_without_allocated_at(*options) end end def self.enable_allocated_at self.send :alias_method, :initialize_without_allocated_at, :initialize self.send :alias_method, :initialize, :initialize_with_allocated_at end end Showing code location
58. # Enabling allocation tracking on every object ObjectSpace.each_object(Class) do |klass| klass.enable_allocated_at end # Building a huge allocation map of all the objects in the system class Object def allocation_map ObjectSpace.each_object.reduce({}) { |hash, object| hash[object.allocated_at] ||= [] hash[object.allocated_at] << object if object.respond_to?(:allocated_at) hash } end end Showing code location
59. Ruby 2.0
60. • Enumerable::Lazy [1, 2, 3].lazy.map { ... }.select bypasses intermediate arrays, albeit is currently very slow Ruby 2.0
61. • Bitmap marking helps to avoid pre-2.0 problem where fork()-CoW is useless because of GC marks Ruby 2.0
62. • caller_locations versus caller is orders of magnitude faster and thus can replace caller in the original allocations_map Ruby 2.0
63. • immediate Floats only on 64-bit systems, speeds up floating point calculations Ruby 2.0
64. is coming in December 2013 Ruby 2.1
65. is coming in December 2013 with generational GC: https://bugs.ruby-lang.org/ issues/8339 Ruby 2.1
66. collects younger objects more often Generational GC
67. Generational GC in action
68. Questions? Thanks