The myth of dynamic language performance

Talk at DevConTLV about how to make your dynamic language implementation faster. What are some basic techniques used and how do they work.

Transcript

1. The myth of dynamic language performance DevConTLV
2. Dirkjan Bussink @dbussink
3. Rubinius A Ruby implementation
5. “PHP too slow for Facebook, builds its own faster version” “Everything you heard about Ruby being slow is not true. It’s about twice as slow as that” “How JavaScript is Slowing Down the Web” 2006
9. Static vs. Dynamic
10. Weak vs. Strong
11. C PHP Java Ruby Static Dynamic Weak Strong
13. The lure of the hash table
14. class MethodTable def initialize @methods = {} end def store_method(name, code) @methods[name] = code end def find_method(name) @methods[name] end end
15. class Address { private String street; private Integer number; private String city; public Address(String street, Integer number, String city) { this.street = street; this.number = number; this.city = city; } }
16. class Address { private HashTable<String, Object> instance_variables; public Address(String street, Integer number, String city) { instance_variables = new HashTable<String, Object>(); instance_variables.put("street", street); instance_variables.put("number", number); instance_variables.put("city", city); } }
17. Hash tables everywhere!
18. Inline caching
19. p = Person.new p.name <receiver_class_id, code>
20. Specific object layout
21. class Address attr_reader :street attr_reader :number attr_reader :city end Address.instance_variable_get("@seen_ivars") => [:@street, :@number, :@city]
22. 0001: push_ivar :@number self[1] Removes hash table lookup
23. Just in time compilation
24. def method1 1 + method2 end def method2 2 + 1 end 100000.times do method1 end ... 0x110ed90e7 mov $0x9, %eax 0x110ed90ec jmp 0x119 ; 0x110ed9129 ... 0x110ed9129 addq $0xa0, %rsp 0x110ed9130 pop %rbp 0x110ed9131 ret 1 + 2 + 1 = 4 FIXNUM(4) = 0x9
25. What about you as a developer?
26. Be nice
27. Write type stable code
28. Small and simple methods
29. Benchmark!
30. Recap
31. Fast dynamic languages are harder, not impossible
32. Writing simple and well factored code is often easy to optimize for a VM
33. ?