How to optimize Ruby internal.

"Ruby 3" has aimed to optimize performance which is one of goals to release. I have made some patches to optimize Ruby internal to realize it.

This talk describes how optimized Ruby internal at Ruby 2.5.

Transcript

1. How to optimize Ruby internal Shizuo Fujita
2. Self • @watson1978 • Ubiregi Inc. • Ruby committer
3. Ruby 3x3 • Ruby 3 need 3 times faster performance than Ruby 2
4. About this talk • How to measure Ruby Internal • Idea to optimize Ruby internal • Optimize
5. How to measure
6. Prepare benchmark code hash1 = {aaa: 12, bbb: 34} hash2 = {ccc: 56, ddd: 78} loop do hash1.merge(hash2) end
7. Mesure $ iprofiler -timeprofiler ./miniruby ~/benchmark.rb hash1 = {aaa: 12, bbb: 34} hash2 = {ccc: 56, ddd: 78} loop do hash1.merge(hash2) end
12. Idea to optimize Ruby internal
13. Method execution... • Method dispatching • Look up constants / methods • Ruby method executing • Implemented in Hash/String/Array/Time...
14. Dispatching Method execution
15. Dispatching Method execution
16. Dispatching Method execution
17. Dispatching Method execution
18. • Focused to reduce method execution time
19. • Focused to reduce method execution time • Remove dispatching in method • Remove redundant allocations
20. Remove dispatching Dispatching Method execution Dispatching Ruby method via rb_funcall()
21. Remove dispatching Dispatching Method execution
22. Remove redundant allocations BMMPDBUJPOT Dispatching Method execution
23. Remove redundant allocations BMMPDBUJPOT BMMPDBUJPOT Dispatching Method execution
24. Optimize
25. Hash#merge • It has used rb_obj_dup() and it’s calling rb_funcall()
26. rb_obj_dup() • It calls Object#initialize_dup via rb_funcall()
27. rb_obj_dup() • It calls Object#initialize_dup via rb_funcall() • Replace rb_obj_dup() to something like rb_ary_dup() to remove redundant Object#initialize_dup
28. Patch for Hash#merge
29. Patch for Hash#merge • Replaced rb_obj_dup() to remove rb_funcall()
30. Hash#merge performance Y Y Y Y Y Y 3VCZ
     3VCZ
    EFW Y Y hash1 = { "a" => 100, "b" => 200 } hash2 = { "b" => 254, "c" => 300 } hash1.merge(hash2)
31. Patch for Time (1)
32. Patch for Time (2)
33. Patch for Time (3)
34. Time methods • Time methods called Ruby methods via rb_funcall()
35. Time methods • Time methods called Ruby methods via rb_funcall() • Added some internal APIs to call method directly
36. Time#- performance Y Y Y Y Y Y 3VCZ
     3VCZ
    EFW Y Y Time.now - Time.at(0)
37. Others
38. Result (2.4.1 vs 2.5.0-dev) "SSBZ Y Y Y Y Y )BTI Y Y Y Y Y 4USJOH Y Y Y Y Y 5JNF Y Y Y Y Y Ubuntu 17.04 gcc version 7.0.1 ruby 2.5.0dev (2017-08-27 trunk 59665) [x86_64-linux]
39. Top 10 4USJOH<OUI> PUIFS  4USJOHJOTFSU QPT
    PUIFS  "SSBZSBTTPD PCK  5JNFTVCTFD  5JNFUP@J  5JNFUW@TFD  )BTIIBT@WBMVF
    OP
    WBMVF  )BTIWBMVF
    OP
    WBMVF  5JNFUP@S  "SSBZNBY O 
40. Worst 10 "SSBZDZDMF O
    \cPCKc
    CMPDL
    ^  "SSBZFBDI@JOEFY
    \cJOEFYc
    ^  4USJOHUP@J  "SSBZBOZ
    \cYc
    CMPDL
    ^  "SSBZSJOEFY WBM
    OPU
    GPVOE  4USJOHMJOFT   5JNFVTFD  "SSBZCTFBSDI@JOEFY
    \cYc
    CMPDL
    ^  4USJOHMJOFT 
    \cMJOFc
    
    ^  )BTI
    MJUFSBM 
41. Worst 10 "SSBZDZDMF O
    \cPCKc
    CMPDL
    ^  "SSBZFBDI@JOEFY
    \cJOEFYc
    ^  4USJOHUP@J  "SSBZBOZ
    \cYc
    CMPDL
    ^  "SSBZSJOEFY WBM
    OPU
    GPVOE  4USJOHMJOFT   5JNFVTFD  "SSBZCTFBSDI@JOEFY
    \cYc
    CMPDL
    ^  4USJOHMJOFT 
    \cMJOFc
    
    ^  )BTI
    MJUFSBM 
42. Ruby 2.5.0-dev 35.7 % slow down
43. 3FHSFTTJPO
    XBT
    pYFE
    CZ
    TIZPVIFJ
44. One more thing…
45. Hash • Hash object need to allocate some heap areas
46. Hash Internal RBasic st_table * int VALUE RHash char char char int st_hash_type * st_index_t st_index_t * st_index_t st_index_t st_table_entry * st_table st_index_t st_index_t st_index_t st_index_t …. st_index_t [] st_table_entry st_table_entry …. st_table_entry st_table_entry []
47. 4 allocations RBasic st_table * int VALUE RHash char char char int st_hash_type * st_index_t st_index_t * st_index_t st_index_t st_table_entry * st_table st_index_t st_index_t st_index_t st_index_t …. st_index_t [] st_table_entry st_table_entry …. st_table_entry st_table_entry []
48. Reused & faster Slow allocating RBasic st_table * int VALUE RHash char char char int st_hash_type * st_index_t st_index_t * st_index_t st_index_t st_table_entry * st_table st_index_t st_index_t st_index_t st_index_t …. st_index_t [] st_table_entry st_table_entry …. st_table_entry st_table_entry []
49. Always allocating RBasic st_table * int VALUE RHash char char char int st_hash_type * st_index_t st_index_t * st_index_t st_index_t st_table_entry * st_table st_index_t st_index_t st_index_t st_index_t …. st_index_t [] st_table_entry st_table_entry …. st_table_entry st_table_entry []
50. …. …. …. …. Before After Concatenate heap areas 2 allocations 1 allocation
51. Hash literal performance Y Y Y Y Y Y Y Before After h = {foo: 12, bar: 34, baz: 56} Caution: This is just prototype
52. vs. Ruby 2.4.1 After Ruby 2.4.1 Base : ruby 2.5.0dev (2017-09-10 trunk 59745) [x86_64-linux] Y Y Y Y Y Y Y Caution: This is just prototype h = {foo: 12, bar: 34, baz: 56}
53. You might learn: • How to measure • Some ways to optimize effectively • A part of current Ruby-dev status
54. Thank you !!