Alexander Dymo - RubyConf 2014 - Ruby Performance Secrets and How to Uncover Them
2 likes657 views
The document discusses Ruby performance optimization techniques. It provides an example of optimizing an inject method call by eliminating unnecessary object allocations. It also analyzes the memory usage of gsub and gsub! string methods through profiling. The key lessons are to use profilers to understand performance bottlenecks, challenge assumptions, and look at Ruby source code and use a debugger when needed to fully understand behavior.
![(gdb) r -e '[1,2,3].inject {}'
Starting program:
/home/gremlin/.rbenv/versions/2.1.4/bin/ruby -e
'[1,2,3].inject {}'
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib64/libthread_db.so.1".
[New Thread 0x7ffff7ff2700 (LWP 3893)]
Breakpoint 1, enum_inject (argc=0, argv=<optimized out>,
obj=93825001586240) at enum.c:640
640 rb_block_call_func *iter = inject_i;
(gdb)](https://image.slidesharecdn.com/adymo-rubyconf-performance-141125184600-conversion-gate02/85/Alexander-Dymo-RubyConf-2014-Ruby-Performance-Secrets-and-How-to-Uncover-Them-29-320.jpg)
![10000.times { [].inject }
20000 extra T_NODE objects
some work for GC](https://image.slidesharecdn.com/adymo-rubyconf-performance-141125184600-conversion-gate02/85/Alexander-Dymo-RubyConf-2014-Ruby-Performance-Secrets-and-How-to-Uncover-Them-44-320.jpg)
Alexander Dymo - RubyConf 2014 - Ruby Performance Secrets and How to Uncover Them
- 1. Ruby Performance Secrets and How to Uncover Them http://www.slideshare.net/adymo/adymo-rubyconf-performance
- 2. Who am I? Alexander Dymo C/C++ since 2000 Ruby/Rails since 2006 Started to optimize back in 2007 Never stopped since then
- 3. Rails Performance: What You Need to Know https://www.airpair.com/ruby-on-rails/performance Make Your Ruby/Rails App Fast: Performance And Memory Profiling Using ruby-prof and Kcachegrind http://www.acunote.com/blog/2008/02/make-your-ruby-rails-applications-fast-performance-and-memory-profiling.html Ruby Performance Tuning http://theprosegarden.com/contents-of-recent-issues/#10-14
- 4. Ruby Performance The first comprehensive book on Ruby Performance I'm 50% done. Beta soon. ruby-performance-book.com
- 6. What do we talk about today? Performance tips Performance best practices
- 7. What do we talk about today? Performance tips Performance best practices How to understand what's wrong How to find your own performance tips/best practices
- 8. In examples
- 9. Example 1
- 10. What can go wrong with this code?
- 11. What can go wrong with this code?
- 12. This was faster
- 15. Smells like...
- 17. Let's check what happens:
- 18. Let's profile memory allocations Need patched ruby rvm reinstall 1.9.3 --patch railsexpress rvm reinstall 2.0.0 --patch railsexpress rvm reinstall 2.1.4 --patch railsexpress
- 19. Let's profile memory allocations Need profiler gem install ruby-prof
- 20. Let's profile memory allocations Need visualization tool Mac: brew install qcachegrind Linux: <your package manager> install kcachegrind Windows: http://sourceforge.net/projects/qcachegrindwin/
- 21. Let's profile memory allocations ruby-prof -p call_tree –mode=allocations before.rb > callgrind.out.before ruby-prof -p call_tree –mode=allocations after.rb > callgrind.out.after kcachegrind callgrind.out.before kcachegrind callgrind.out.after
- 25. static VALUE enum_inject(int argc, VALUE *argv, VALUE obj) { NODE *memo; VALUE init, op; rb_block_call_func *iter = inject_i; … memo = NEW_MEMO(init, argc, op); rb_block_call(obj, id_each, 0, 0, iter, (VALUE)memo); return memo->u1.value; }
- 26. > gdb `rbenv which ruby` GNU gdb (GDB) SUSE (7.5.1-2.5.1) Reading symbols from /home/gremlin/.rbenv/versions/2.1.4/bin/ruby...done. (gdb)
- 27. (gdb) l enum_inject 632 * longest #=> "sheep" 633 * 634 */ 635 static VALUE 636 enum_inject(int argc, VALUE *argv, VALUE obj) 637 { 638 NODE *memo; 639 VALUE init, op; 640 rb_block_call_func *iter = inject_i; 641 ID id; (gdb)
- 28. 636 enum_inject(int argc, VALUE *argv, VALUE obj) 637 { 638 NODE *memo; 639 VALUE init, op; 640 rb_block_call_func *iter = inject_i; 641 ID id; (gdb) b 638 Breakpoint 1 at 0x1cbc0a: file enum.c, line 638. (gdb)
- 29. (gdb) r -e '[1,2,3].inject {}' Starting program: /home/gremlin/.rbenv/versions/2.1.4/bin/ruby -e '[1,2,3].inject {}' [Thread debugging using libthread_db enabled] Using host libthread_db library "/lib64/libthread_db.so.1". [New Thread 0x7ffff7ff2700 (LWP 3893)] Breakpoint 1, enum_inject (argc=0, argv=<optimized out>, obj=93825001586240) at enum.c:640 640 rb_block_call_func *iter = inject_i; (gdb)
- 30. 640 rb_block_call_func *iter = inject_i; (gdb) n 665 memo = NEW_MEMO(init, argc, op); (gdb)
- 31. 640 rb_block_call_func *iter = inject_i; (gdb) n 665 memo = NEW_MEMO(init, argc, op); (gdb) n 666 rb_block_call(obj, id_each, 0, 0, iter, (VALUE)memo); (gdb)
- 32. 640 rb_block_call_func *iter = inject_i; (gdb) n 665 memo = NEW_MEMO(init, argc, op); (gdb) n 666 rb_block_call(obj, id_each, 0, 0, iter, (VALUE)memo); (gdb) s rb_block_call (obj=93825001586240, mid=1456, argc=0, argv=0x0, bl_proc=0x555555722460 <inject_i>, data2=93825001586200) at vm_eval.c:1142 1142 { (gdb)
- 33. 640 rb_block_call_func *iter = inject_i; (gdb) n 665 memo = NEW_MEMO(init, argc, op); (gdb) n 666 rb_block_call(obj, id_each, 0, 0, iter, (VALUE)memo); (gdb) s rb_block_call (obj=93825001586240, mid=1456, argc=0, argv=0x0, bl_proc=0x555555722460 <inject_i>, data2=93825001586200) at vm_eval.c:1142 1142 { (gdb) s 1145 arg.obj = obj; (gdb)
- 34. 640 rb_block_call_func *iter = inject_i; (gdb) n 665 memo = NEW_MEMO(init, argc, op); (gdb) n 666 rb_block_call(obj, id_each, 0, 0, iter, (VALUE)memo); (gdb) s rb_block_call (obj=93825001586240, mid=1456, argc=0, argv=0x0, bl_proc=0x555555722460 <inject_i>, data2=93825001586200) at vm_eval.c:1142 1142 { (gdb) s 1145 arg.obj = obj; (gdb) s 1146 arg.mid = mid; (gdb)
- 35. 640 rb_block_call_func *iter = inject_i; (gdb) n 665 memo = NEW_MEMO(init, argc, op); (gdb) n 666 rb_block_call(obj, id_each, 0, 0, iter, (VALUE)memo); (gdb) s rb_block_call (obj=93825001586240, mid=1456, argc=0, argv=0x0, bl_proc=0x555555722460 <inject_i>, data2=93825001586200) at vm_eval.c:1142 1142 { (gdb) s 1145 arg.obj = obj; (gdb) s 1146 arg.mid = mid; (gdb) s 1147 arg.argc = argc; (gdb)
- 36. (gdb) s 1147 arg.argc = argc; (gdb) s 1148 arg.argv = argv; (gdb)
- 37. (gdb) s 1147 arg.argc = argc; (gdb) s 1148 arg.argv = argv; (gdb) s 1149 return rb_iterate(iterate_method, (VALUE)&arg, bl_proc, data2); (gdb)
- 38. (gdb) s 1147 arg.argc = argc; (gdb) s 1148 arg.argv = argv; (gdb) s 1149 return rb_iterate(iterate_method, (VALUE)&arg, bl_proc, data2); (gdb) s rb_iterate (it_proc=it_proc@entry=0x5555556c0790 <iterate_method>, data1=data1@entry=140737488340304, bl_proc=0x555555722460 <inject_i>, data2=93825001586200) at vm_eval.c:1054 1054 { (gdb)
- 39. (gdb) s 1147 arg.argc = argc; (gdb) s 1148 arg.argv = argv; (gdb) s 1149 return rb_iterate(iterate_method, (VALUE)&arg, bl_proc, data2); (gdb) s rb_iterate (it_proc=it_proc@entry=0x5555556c0790 <iterate_method>, data1=data1@entry=140737488340304, bl_proc=0x555555722460 <inject_i>, data2=93825001586200) at vm_eval.c:1054 1054 { (gdb) s 1057 NODE *node = NEW_IFUNC(bl_proc, data2); (gdb)
- 40. (gdb) s 1147 arg.argc = argc; (gdb) s 1148 arg.argv = argv; (gdb) s 1149 return rb_iterate(iterate_method, (VALUE)&arg, bl_proc, data2); (gdb) s rb_iterate (it_proc=it_proc@entry=0x5555556c0790 <iterate_method>, data1=data1@entry=140737488340304, bl_proc=0x555555722460 <inject_i>, data2=93825001586200) at vm_eval.c:1054 1054 { (gdb) s 1057 NODE *node = NEW_IFUNC(bl_proc, data2); (gdb)
- 41. static VALUE enum_inject(int argc, VALUE *argv, VALUE obj) { NODE *memo; VALUE init, op; rb_block_call_func *iter = inject_i; … memo = NEW_MEMO(init, argc, op); rb_block_call(obj, id_each, 0, 0, iter, (VALUE)memo); return memo->u1.value; }
- 42. VALUE rb_block_call(…) { … return rb_iterate(iterate_method, (VALUE)&arg, bl_proc, data2); } VALUE rb_iterate(…) { int state; volatile VALUE retval = Qnil; NODE *node = NEW_IFUNC(bl_proc, data2); … }
- 43. 2 T_NODE's per inject() call
- 44. 10000.times { [].inject } 20000 extra T_NODE objects some work for GC
- 46. Ruby Performance More in my book ruby-performance-book.com
- 47. Lessons learned: 1. use profiler to understand why your code is slow 2. use C debugger to understand Ruby behavior
- 48. Example 2
- 49. What's the difference? str = 'a'*1024*1024*10 str = str.gsub('a', 'b') str = 'a'*1024*1024*10 str.gsub!('a', 'b')
- 50. str = 'a'*1024*1024*10 str = str.gsub('a', 'b') str = 'a'*1024*1024*10 str.gsub!('a', 'b') replaces 'a' with 'b' creates a new object reuses "str" name replaces 'a' with 'b' changes the original
- 51. Supposedly
- 52. Let's profile memory usage ruby-prof -p call_tree –mode=memory after.rb > callgrind.out.after kcachegrind callgrind.out.after
- 54. So, gsub! doesn't save any memory
- 55. So, gsub! doesn't save any memory … except one slot on Ruby heap
- 56. So, gsub! doesn't save any memory except one slot on Ruby heap … which is 40 bytes
- 57. Not all bang! functions are the same str = 'a'*1024*1024*10 str.downcase! ruby-prof -p call_tree –mode=memory downcase.rb > callgrind.out.downcase kcachegrind callgrind.out.downcase
- 59. Lessons learned: 1. profile memory 2. challenge all tips/tricks/best practices
- 60. Conclusions 1. Don't guess. Profile. 2. Guess. Profile. 3. Profile not only CPU, but Memory. 4. Look at the source, use GDB if not enlightened. 5. Challenge all tips/tricks. Understand instead.
- 61. Big thanks to:
- 62. Ruby Performance ruby-performance-book.com airpair.me/adymo @alexander_dymo