Presentation on LMAX disruptor @ java.il 22/12/2015

1. The LMAX Disruptor
Amir Langer - Outbrain (… and ex-LMAX)
Pictures courtesy of Trisha Gee (… another ex-LMAX)

2. Disruptor
High Performance Inter-Thread Messaging Library

3. Disruptor
• Why?
• What?
• Write flow (Event publishing)
• Read flow (Event processing)
• Behind the scenes

5. A Typical LMAX Service

6. A Typical LMAX Service

7. Why queues suck?
(as inter-thread messaging mechanism)

8. Linked List not much better
Usually worse

9. Cost of Contention
Increment a long counter 500M times
One Thread 300ms

10. Cost of Contention
Increment a long counter 500M times
One Thread 300ms
One Thread (Volatile) 4700ms x 15

11. Cost of Contention
Increment a long counter 500M times
One Thread 300ms
One Thread (Volatile) 4700ms x 15
One Thread (Atomic) 5700ms x 19

12. Cost of Contention
Increment a long counter 500M times
One Thread 300ms
One Thread (Volatile) 4700ms x 15
One Thread (Atomic) 5700ms x 19
One Thread (Lock) 10000ms x 33

13. Cost of Contention
Increment a long counter 500M times
One Thread 300ms
One Thread (Volatile) 4700ms x 15
One Thread (Atomic) 5700ms x 19
One Thread (Lock) 10000ms x 33
Two Threads (Atomic) 30000ms x 100

14. Cost of Contention
Increment a long counter 500M times
One Thread 300ms
One Thread (Volatile) 4700ms x 15
One Thread (Atomic) 5700ms x 19
One Thread (Lock) 10000ms x 33
Two Threads (Atomic) 30000ms x 100
Two Threads (Lock) 224000ms x 746
That’s almost
4 minutes!!!

15. Disruptor

16. Building Blocks
• RingBuffer
• EventFactory
• WaitStrategy
• Publisher
• EventTranslator
• EventProcessor
• Sequence & SequenceBarrier

17. • RingBuffer
• EventFactory
• WaitStrategy
• Publisher
• EventTranslator
• EventProcessor
• Sequence & SequenceBarrier
Building Blocks

18. • Events are mutable instances of T
• Sequence denotes last available event
• Publisher will publish to the next slot
• EventProcessor will read up to including sequence

19. • RingBuffer
• EventFactory
• WaitStrategy
• Publisher
• EventTranslator
• EventProcessor
• Sequence & SequenceBarrier
Building Blocks

27. • SequenceBarrier
• GatingSequence

28. • RingBuffer
• EventFactory
• WaitStrategy
• Publisher
• EventTranslator
• EventProcessor
• Sequence & SequenceBarrier
Building Blocks

43. Single Writer
Principle

44. Memory Barrier

45. Memory Barrier

46. Memory Barrier
• Load Barrier - invalidates CPU cache and forces all load
instructions after the barrier to happen after the barrier
• Store Barrier - forces all store instructions prior to the barrier to
happen before the barrier and have the store buffers flushed to
cache for the CPU on which it is issued.
a type of instruction that causes a central processing unit (CPU) to enforce an
ordering constraint on memory operations issued before and after the barrier
instruction

47. in Java
volatile - a volatile field has a store barrier inserted
after a write to it and a load barrier inserted before a
read of it.

49. By Sticking to the Single Writer
Principle we can use only volatile
counters to synchronize between
the producer and eventProcessor

50. What about
Multiple Producers?
Multiple Consumers?

51. Why queues suck? (part 2)

52. Zero Garbage
Oh, and its all Mutable!!!

53. Why queues suck? (Part 3)

54. Batching Effect
Without any contention cost during it

55. Going Parallel

56. How Fast Is It - Latency
ArrayBlockingQueue Disruptor
Min 145 29
Mean 32,757 52
99 Percentile 2,097,152 128
99.99 Percentile 4,194,304 8,192
Max 5,069,086 175,567

57. How Fast Is It - Throughput
0.0
7500000.0
15000000.0
22500000.0
30000000.0
Unicast Diamond
ABQ Disruptor

58. When not to use it?
• Events are not deltas but snapshots so that only the
latest needs to be dealt with
• No way of separating your domain logic to multiple
instances (You’re locked to either having shared
state between instances (DB) or one instance with
multiple threads)

59. Thanks for listening
Questions?
• https://github.com/LMAX-Exchange/disruptor
• http://mechanical-sympathy.blogspot.co.uk/
• https://www.cs.umd.edu/~pugh/java/memoryModel/